Citation
Autonomous methods for vessel reconstruction and quantification of vascular flow

Material Information

Title:
Autonomous methods for vessel reconstruction and quantification of vascular flow
Creator:
Yunker, Bryan E. ( author )
Place of Publication:
Denver, CO
Publisher:
University of Colorado Denver
Publication Date:
Language:
English
Physical Description:
1 electronic file (223 pages). : ;

Thesis/Dissertation Information

Degree:
Doctorate ( Doctor of Philosophy)
Degree Grantor:
University of Colorado Denver
Degree Divisions:
Department of Bioengineering, CU Denver
Degree Disciplines:
Bioengineering

Subjects

Subjects / Keywords:
Blood vessls -- Tumors ( lcsh )
Hemostasis ( lcsh )
Tumors ( lcsh )
Microcirculation ( lcsh )
Genre:
bibliography ( marcgt )
theses ( marcgt )
non-fiction ( marcgt )

Notes

Review:
Introduction: The success rate for the Radiofrequency (RF) ablation of hepatic tumors is limited by local blood flow carrying away the heat applied during treatment, which can prevent some tumor cells from reaching the necrosis (cell death) temperature of 55 °C. As previous studies [1] suggest that local flow may be used as a predictor of ablation size and device performance, the focus of this study explores a method of determining local flow by reconstructing hepatic vasculature structure from 3D ultrasound. With the hypothesis that the outer surface of 3D Doppler data can accurately represent the inner surface (lumen) of a vessel, two software applications (ADP and VAS) were created as platforms for developing and testing algorithms capable of meeting a clinical Time-to-Results (TTR) goal of 10 seconds for local flow estimation. Methods: A portal vein structure was acquired for use as a pattern for flow phantom construction and CFD flow simulations. Industrial casting materials were imaged with ultrasound, MRI and CT, and then used to fabricate flow phantoms. The flows through the portal phantom branches were measured over a range of flow rates. A Computational Fluid Dynamics simulation was created and compared to the measured phantom branch flows. Algorithms capable of estimating local flow were developed and evaluated using simple tubing flow phantoms. Results/Discussion: Each of the materials tested for phantoms showed Results/Discussion: Each of the materials tested for phantoms showed useful characteristics in medical imaging applications. Two portal vein flow phantoms were successfully fabricated, one as a solid silicone block, and the other as a thin walled latex tube structure. The portal phantom branch flow measurements and the CFD model estimates compared within 17%. Color Doppler surfaces within the phantom lumen could not be obtained with the water based test fluids used, however: testing with Doppler surfaces ~2x larger than reference provided random diameter estimate deviations under 9%, angle corrected velocities within 15% of reference, and net flows within 15% of reference. The segmentation, centerline, diameter, area, angle correction, and flow algorithms functioned correctly and within the TTR. Conclusions: The rapid prototyping fabrication techniques and materials employed resulted in usable anatomical flow phantoms. CFD modeling successfully predicted the performance of the anatomical flow phantoms. The algorithms developed for local flow assessment are ready for accuracy testing with a more echogenic test fluid (blood). The diameter/area algorithm needs to be improved to handle non-circular vascular structure. The error performance for all algorithms needs to be improved.
Thesis:
Thesis (Ph.D.)--University of Colorado Denver. Bioengineering
Bibliography:
Includes bibliographic resources.
General Note:
Department of Bioengineering
Statement of Responsibility:
by Bryan E. Yunker.

Record Information

Source Institution:
University of Colorado Denver
Holding Location:
|Auraria Library
Rights Management:
All applicable rights reserved by the source institution and holding location.
Resource Identifier:
902811541 ( OCLC )
ocn902811541

Downloads

This item has the following downloads:


Full Text
AUTONOMOUS METHODS FOR VESSEL RECONSTRUCTION AND
QUANTIFICATION OF VASCULAR FLOW by
BRYAN E. YUNKER Certificate, University of Colorado, 2009 Bachelor of Science, Washington State University, 1981
A thesis submitted to the Faculty of the Graduate School of the University of Colorado in partial fulfillment of the requirements for the degree of Doctor of Philosophy Bioengineering Program
2014


2014
BRYAN E. YUNKER ALL RIGHTS RESERVED
11


This thesis for the Doctor of Philosophy degree by Bryan E. Yunker has been approved for the Bioengineering Program by
Robin Shandas, Chair Kendall S. Hunter, Advisor Gerald D. Dodd Yusheng Feng S. James Chen
Samuel Chang


Yunker, Bryan E. (Ph.D., Bioengineering)
Autonomous Methods for Vessel Reconstruction and Quantification of Vascular Flow
Thesis directed by Assistant Professor Kendall S. Hunter.
ABSTRACT
Introduction: The success rate for the Radiofrequency (RF) ablation of hepatic tumors is limited by local blood flow carrying away the heat applied during treatment, which can prevent some tumor cells from reaching the necrosis (cell death) temperature of 55 C.
As previous studies [1] suggest that local flow may be used as a predictor of ablation size and device performance, the focus of this study explores a method of determining local flow by reconstructing hepatic vasculature structure from 3D ultrasound. With the hypothesis that the outer surface of 3D Doppler data can accurately represent the inner surface (lumen) of a vessel, two software applications (ADP and VAS) were created as platforms for developing and testing algorithms capable of meeting a clinical Time-to-Results (TTR) goal of 10 seconds for local flow estimation. Methods: A portal vein structure was acquired for use as a pattern for flow phantom construction and CFD flow simulations. Industrial casting materials were imaged with ultrasound, MRI and CT, and then used to fabricate flow phantoms. The flows through the portal phantom branches were measured over a range of flow rates. A Computational Fluid Dynamics simulation was created and compared to the measured phantom branch flows. Algorithms capable of estimating local flow were developed and evaluated using simple tubing flow phantoms. Results/Discussion: Each of the materials tested for phantoms showed useful
IV


characteristics in medical imaging applications. Two portal vein flow phantoms were successfully fabricated, one as a solid silicone block, and the other as a thin walled latex tube structure. The portal phantom branch flow measurements and the CFD model estimates compared within 17%. Color Doppler surfaces within the phantom lumen could not be obtained with the water based test fluids used, however: testing with Doppler surfaces ~2x larger than reference provided random diameter estimate deviations under 9%, angle corrected velocities within 15% of reference, and net flows within 15% of reference. The segmentation, centerline, diameter, area, angle correction, and flow algorithms functioned correctly and within the TTR. Conclusions: The rapid prototyping fabrication techniques and materials employed resulted in usable anatomical flow phantoms. CFD modeling successfully predicted the performance of the anatomical flow phantoms. The algorithms developed for local flow assessment are ready for accuracy testing with a more echogenic test fluid (blood). The diameter/area algorithm needs to be improved to handle non-circular vascular structure. The error performance for all algorithms needs to be improved.
The form and content of this abstract are approved. I recommend its publication.
Approved: Kendall S. Hunter
v


DEDICATION
I dedicate this work to my parents, grandparents, family, friends, committee, and colleagues for their unwavering support and patience.


ACKNOWLEDGMENTS
I would like to thank the following individuals for their support, guidance, and teachings:
Committee Members: Kendall S. Hunter, Gerald D. Dodd, Yusheng Feng,
Robin Shandas, S. James Chen, and Samuel Chang Fellow Authors: Dietmar Cordes, Ann L. Scherzinger Liver Lab Manager: Anthony Lanctot Bioengineering Instructor and Lab Manager: Craig Lanning Lab Assistants: Bryan Rech, Jennifer Wagner and Luciano Mazzaro Professors: Gary Fullerton, Emily Gibson, Mark Golkowski,
William Hendee, Willem Schreiider, Mark Brown Fellow Students: Arati Gurung, Shawna Burgett, Steve Humphries, Mike Zimkowski Funding Sources: NIH/NCI1F31CA171620-01A1, Colorado Translational Research Imaging Center (C-TRIC); NSF 0932339; University of Colorado, SOM, Department of Radiology; NIH T32HL072738; K25 HL094749; K24081506; R01HL114753; and NHLBIK25-094749 Equipment: Ged Harrison, Philips Healthcare, Bothell, WA
Vll


TABLE OF CONTENTS
CHAPTER
I. INTRODUCTION...........................................................1
Background and Significance.........................................1
Disease Impact....................................................1
Disease Treatment.................................................2
Research Objective................................................8
Estimating Local Blood Velocity/Flow..............................9
Background Studies.................................................12
Hepatic Physiology...............................................12
Hepatic Flow Rates...............................................22
Vascular Segmentation............................................32
Clinical Requirements............................................41
II. RESEARCH..............................................................47
Specific Aims......................................................47
Aim 1 Fabricate and Model a Portal Vein Flow Phantom...........47
Aim 2 Obtain an ROI from a 3D Image Volume from Clinician Input.47
Aim 3 Autonomously Segment Vasculature and Compute Centerline...47
Aim 4 Autonomously Compute Segmented Vessel Diameters and Cross-
Sectional Areas...........................................48
viii


Aim 5 Autonomously Compute Angle Corrected Velocities for Segmented
Vessels....................................................48
Aim 6 Autonomously Compute Volumetric Flow for Segmented Vessels... 48
Materials and Methods.............................................49
Aim 1 Fabricate and Model a Portal Vein Flow Phantom............49
Aim 2 Obtain an ROI from a 3D Image Volume from Clinician Input.57
Aim 3 Autonomously Segment Vasculature and Compute Centerline...68
Aim 4 Autonomously Compute Segmented Vessel Diameters and Cross-
Sectional Areas............................................75
Aim 5 Autonomously Compute Angle Corrected Velocities for Segmented
Vessels....................................................77
Aim 6 Autonomously Compute Volumetric Flow for Segmented Vessels... 80
Results and Discussion............................................81
Aim 1 Fabricate and Model a Portal Vein Flow Phantom............81
Aim 2 Obtain an ROI from a 3D Image Volume from Clinician Input.91
Aim 3 Autonomously Segment Vasculature and Compute Centerline...93
Aim 4 Autonomously Computed Segmented Vessel Diameters and Cross-
Sectional Areas...........................................104
Aim 5 Autonomously Compute Angle Corrected Velocities for Segmented
Vessels...................................................117
Aim 6 Autonomously Compute Volumetric Flow for Segmented Vessels. 137
III. FULL TEXT OF AIM 1 PHANTOM MATERIALS (PUBLISHED).....................142
IV. FULL TEXT OF AIM 1 PHANTOM FABRICATION (PUBLISHED).................169
V. CONCLUSIONS............................................................187
IX


VI. FUTURE RESEARCH....................................191
REFERENCES.........................................197
x


LIST OF TABLES
TABLE
1.1 Normal Portal Vein Size, Flow, and Pressure Characteristics..................22
1.2 Blood Properties.............................................................26
1.3 Thermal Characteristics of Blood and Liver...................................27
1.4 Hepatic Ultrasound Characteristics...........................................31
1.5 Summary of Vascular Segmentation Approaches..................................37
II. 1 Centerline Error from Surface Noise.......................................103
11.2 Diameter Error Statistics..................................................106
11.3 Diameter Errors from Centerline............................................107
11.4 Area Errors from Diameter Error............................................117
11.5 Uncorrected Velocity Statistics............................................124
11.6 Reynolds Numbers for Tested Images.........................................125
11.7 Straight Tube Velocity Errors (S0324-65B)..................................129
11.8 Curved Tube Velocity Errors (S0324-65R)....................................133
xi


II.9 Rough Surface Velocity Errors (S0313-70B)................................137
II. 10 Net Flow Velocity Errors...............................................141
xii


LIST OF FIGURES
FIGURE
1.1 Layout of RF Ablation Procedure.............................................3
1.2 RF Probe Tip Temperature Gradient...........................................4
1.3 RF Ablation Progression.....................................................4
1.4 Thermal Injury..............................................................5
1.5 Ablation Volume Vs Flow.....................................................7
1.6 Ablation Volume Vs VFI......................................................8
1.7 Typical Doppler Color Ultrasound Image of Hepatic Veins....................10
1.8 Location of the Liver in the Circulatory System............................12
1.9 GI Vasculature Feeding the Portal Vein.....................................13
1.10 Vascular Anatomy of the Liver............................................14
1.11 2D Image of Portal Vein and Hepatic Artery...............................15
1.12 Sections of the Liver....................................................16
. 17 xiii
1.13 Hepatic Lobule Structure


1.14 Hepatic Lobules Flows.......................................................18
1.15 Lobule Cross-section........................................................19
1.16 Detail of the Blood Flow through the Hepatic Sinusoids......................19
1.17 Hepatocyte Cellular Structure...............................................21
1.18 Operating Inlet Pressure of the Liver.......................................23
1.19 Diagram of the Hepatic Flows and Pressures..................................24
F20 Viscoelastic Properties of Blood at 2 Hz and 22 C............................26
F21 Cirrhosis.....................................................................30
1.22 Geometry of Doppler.........................................................31
F23 Example of Manual Angle Correction............................................32
F24 Hepatic Branching ..........................................................33
F25 Arterial Phase CT...........................................................39
F26 Arterial Phase Hepatic Artery...............................................39
F27 Venous Phase CT.............................................................40
F28 Portal Vein...................................................................40
IF 1 Solid Silicone Body Portal Phantom and Test Setup........................50
xiv


11.2 Thin Walled Latex Portal Phantom and Test Setup......................51
11.3 Thin Walled Latex Portal Phantom Test Setup..........................52
11.4 Thin Tube Phantoms....................................................54
11.5 BW and Doppler Slices (BW Mode)......................................58
11.6 Geometry of Doppler (False Color Mode)...............................59
11.7 Red and Blue Doppler (False Color)...................................60
II. 8 Histograms of Red and Blue Doppler..................................61
II.9 Linear Color Mapping of BW and Doppler...............................61
II. 10 VAS Clinical Application Graphical User Interface................62
II. 11 White BW Scan Envelope Overlaid with Red and Blue Doppler..........64
11.12 Raw BW Data Volume (Semi-transparent)................................65
II. 13 Red and Blue Doppler Data in Raw Mode...............................65
11.14 Red and Blue Doppler shown in RawTC (True Color) Mode................67
11.15 Phantom Test Samples.................................................81
II. 16 Solid Silicone Body Portal Vein Phantom............................82
II. 17 Thin Walled Dipped Latex Portal Vein Phantom.......................83
xv


II. 18 Solid Body Silicone Portal Vein Phantom..................................84
II. 19 Thin Walled Latex Portal Vein Phantom....................................85
11.20 Portal Inlet Flow Vs Pressure............................................86
11.21 Portal Branch Flow Rates.................................................87
11.22 Major Portal Branch Flow Rates...........................................88
11.23 Velocity Profile Simulation of Phantom Pattern Mesh......................89
11.24 Comparison of CFD Estimates with Phantom Measurements....................90
11.25 ROI Selection Box (Lateral View).........................................91
11.26 ROI Selection Box (Elevation View).......................................92
11.27 ROI Selection Completed..................................................93
11.28 Edge, Center, and Elliptical Approximations of Vessel Cross-section......94
11.29 Gaussian Smoothing of Doppler Data.......................................94
11.30 Outline and Center Computed for Vessel Cross-section.....................95
11.31 Typical Vessel Overlap and Proximity......................................96
11.32 Segmented Point Cloud of Hepatic Capsule and Portal/Hepatic Veins........96
11.33 Centerline Approximation.................................................97
xvi


11.34 Multi-Vessel ROI Capture................................................98
11.35 Surface Noise in Color Doppler Point Clouds.............................99
11.36 Long Curved Vessel Test (S0324-65R)....................................101
11.37 Centerline Direction Error Due to Surface Noise........................102
11.38 Color Doppler Cross-section Capture (S0324-65B-3)......................104
11.39 Color Doppler Cross-section Capture (S0313-70B-5)......................104
11.40 Color Doppler Diameter Measurements....................................105
11.41 Color Doppler Diameter Error Eccentricity (S0313-70)................ 108
11.42 Color Doppler Diameter Error Pattern (S0320-34)..................... 110
11.43 Surface Detection Threshold Set to 1/128 and 25/128................... 112
11.44 End Tapering of Color Doppler Surfaces (S0324-65BR)...................114
11.45 Power Doppler Diameter Measurement Tubing (S0407-41)................ 114
11.46 BW Image Envelope and Computed Transducer Origins......................118
11.47 Transducer Beam Lines..................................................119
11.48 Comparison of Mass/Time and Manual Angle Correction....................120
11.49 Uncorrected Velocities.................................................122
xvii


11.50 Location of Velocity Samples...........................................123
11.51 Manual Angle Correction and Color Scale (Cart).........................126
11.52 Manual Angle Correction and Color Scale (VAS)..........................126
11.53 Doppler Velocity True Color Variation (S0324-65)...................... 127
11.54 Straight Tube Velocity Errors (S0324-65B)..............................128
11.55 Straight Tube Blue Top View (S0324-65B)..............................130
11.56 Curved Tube Top View (S0324-65R).....................................132
11.57 Curved Tube Velocity Errors (S0324-65R)................................133
11.58 Series Phantom All Tubes.............................................135
11.59 Rough Surface Capture Top View (S0313-70B-5).........................136
11.60 Rough Surface Velocity Errors (S0313-70B)..............................136
11.61 Net Flow Phantom, Straight-Curved Tubing (S0324-65)................... 138
11.62 Net Flow Errors........................................................140
VI. 1 Extracted Vasculature Registered to Patient Physiology.................194
VI.2 Virtual Reality Demo....................................................195
xviii


LIST OF EQUATIONS
EQUATION
1.1 Velocity Flow Index (VFI)...................................................7
1.2 Flow from Area and Velocity.................................................9
1.3 Hepatic Resistance by Windkessel............................................28
1.4 Resistive Index for Hepatic Artery..........................................28
1.5 Portal Vein Reynolds Number..................................................29
II. 1 Mean Mass to Volume Flow Conversion.......................................53
11.2 Velocity to Volume Flow Conversion.........................................56
11.3 Voxels in the Vessel Cross-sectional Plane.................................76
11.4 Area of Vessel Cross-section................................................77
11.5 Cross-sectional Mean Velocity...............................................78
11.6 Angle Between Insonation Beam and Centerline Segment.......................79
11.7 Centerline Error from Noise................................................103
11.8 Centerline Error...........................................................107
11.9 Elliptical Area............................................................116
xix


II. 10 Test for Laminar Vs Turbulent Flow
125
xx


LIST OF ABBREVIATIONS
ADP Algorithm Development Platform BW Black and White CT Computed Tomography GUI Graphical User Interface HCC Hepatocellular Carcinoma MRI Magnetic Resonance Imaging RF Radiofrequency RI Resistive Index ROI Region of Interest TTR Time-To-Results VAS Vascular Analysis System VFI Vascularization Flow Index


CHAPTER I
INTRODUCTION
Background and Significance Disease Impact
The American Cancer Society (ACS) hepatocellular carcinoma (HCC) and bile duct cancer estimates for 2014 are 33,190 new diagnoses and 23,000 deaths in the United States, where men will lead women 3:1 in diagnosis and 2:1 in deaths [2], According to the NCI SEER report for 2003-2009, the survival rate from hepatic tumors that have spread to other parts of the body is 2%, and, the survival rate from local tumors is 28% [3], Worldwide, the ACS states an average of 700,000 diagnoses and 600,000 deaths [2], and Cancer Research UK states 745,000 deaths in 2012 [4],
In addition to HCC diagnoses and deaths, over 50% of the individuals diagnosed with colorectal cancer will develop liver metastases, comprising 150,000 diagnoses and 50,000 deaths as of 2008 in the United States [5], Colorectal cancer is the third most common type of cancer at 1.4 million cases annually worldwide, suggesting 75,000 more cases of liver cancer in the United States and 700,000 more cases worldwide beyond primary HCC statistics.
In 2013, liver cancer was the fifth most common cause of cancer death in males in the US across all ages [6],
1


Disease Treatment
Systemic chemotherapy is an option for tumor treatment but cures are rare. Surgery offers better survival rates and a longer disease-free period, however, this applies only to a very limited patient population, and, there can be significant postoperative morbidity and high tumor recurrence. Ablation Therapies can be used to destroy the viability of non-resectable tumor cells including Cryogenic Ablation with liquid nitrogen, Chemical Ablation with a toxin such as Ethanol, and Thermal Ablation with Radio Frequency or Microwave energy.
Therapeutic Focus
RF Ablation (RFA) was chosen for this study as it is a less invasive therapeutic technique that can be used repeatedly to treat inevitable tumor recurrences and it is a therapy performed by the clinical sponsor of this research. There are a number of benefits of RFA [7], including local cure rates [complete ablation] can reach 83% for tumors (<3 cm) with minimal side effects, a major complication rate of only 1-3%, and, patients can go home the same day or after an overnight hospital stay. Where RF ablation has been shown to be effective for the treatment of small (< 3 cm) malignant hepatic tumors, its utility in treating larger tumors is quite variable [8],
2


RFA Technology
RF Ablation uses an alternating electric current generator operated at 500 kHz, a needle electrode, and large adhesive ground pads placed under the patient as depicted in Fig. FI.
Figure 1.1 Layout of RF Ablation Procedure [9]
When the RF generator is turned on, the concentrated alternating electric current around the needle electrode causes ionic agitation and frictional heat in the surrounding tissue. An elevation in local temperature higher than 55 C causes coagulation necrosis and cell death in tissue and tumors. The thermal injury created by a straight rod probe is elliptical in shape. The probe tip must operate at temperatures over 120 C in order to heat to the entire tumor volume to the necrosis temperature as shown in Fig. 1.2.
3


Figure 1.2 RF Probe Tip Temperature Gradient
The RF ablation process is shown in Fig. F3 where the left pane shows the untreated tumor. The middle image shows a thermal injury much larger than the tumor, margin required to ensure that all of the tumor cells reach necrosis temperature.
http://radiology.ucla.edu
Figure 1.3 RF Ablation Progression


The right image of 1.3 shows the post operative scar. As the tumor may be larger than the elliptical injury created by a rod probe, the tumor must be treated by either many thermal injuries delivered at intervals within the tumor volume [10] with a single tine probe, or, one thermal injury with a multi-tined probe which would cover a larger volume [9] as shown in Fig. 1.4.
http ;//radiology.ucla. edu/body. cfm?id=l 62
Figure 1.4 Thermal Injury
Technical success of the ablation procedure is determined by whether the entirety of the targeted tumor as well as a circumferential 5-10 mm margin of normal hepatic parenchyma is included in the ablation zone. If a portion of the obvious or microscopic tumor in the immediately adjacent hepatic parenchyma is not included in the ablation zone, local tumor recurrence is inevitable.
5


RF Device Performance
There are several manufacturers of RF ablation devices and all produce ablations that vary in size between 3-7 cm in diameter. This variation has been unpredictable to date and accounts for the success in treating tumors less than 3 cm in diameter and the high failure rates seen in tumors greater than 3 cm in diameter [11, 12], The dominant factor controlling the volume of coagulation necrosis produced by an RF ablation device is the volume of hepatic blood flow in the region of the tumor [12-14], Prior work has shown that the size of the ablation created in a liver by an RF ablation device decreases with increased local blood flow [15-24], In work performed by Chang [1], fresh bovine livers were obtained from a local slaughterhouse and each liver was perfused by its own blood with heparin added to prevent coagulation. RF ablations were performed at several sites within the liver microvasculature at several blood flow rates and the ablation volume was verified by tissue sectioning and MRI. The blood flows were normalized for liver mass and the data plotted in Fig. 1.5 shows a correlation between RF ablation Volume and normalized Blood Flow.
6


30
V -0.556x 31.574
10 3 35 35
Normalized Blood Flow (mi/min/iOOg tissue)
so
Figure 1.5 Ablation Volume Vs Flow [1] [25]
An additional relationship was investigated using Color Doppler imagery of the ablation sites and a Volume Flow Index (VFI) computed with Philips QLab software based on Eq. 1.1.
VFI
Ici Ic E
C = Number of colored pixels in ROI E = Total number of pixels in ROI
Equation 1.1 Velocity Flow Index (VFI)
7


A modest correlation was found between ablation size and the computed VFI as shown in Fig. 16. however, the errors are too large to be clinically useful without additional information.
Figure 1.6 Ablation Volume Vs VFI [1]
Research Objective
To increase the success rates stated by Lu [7] and to expand the number of eligible patients with larger tumors, the size of the ablation zone that will be produced by a specific ablation device in a given patient must be known in advance of the treatment. With this knowledge, the clinician can develop an appropriate treatment plan that
8


includes the position and number of ablations that will need to be performed to produce an ablation zone of sufficient size to eradicate a tumor without excessive margins.
As the relationships shown in Figs. 1.5-6 suggest that ablation size prediction and improved surgical success are dependent on quantification of local blood flow, the research presented in this dissertation attempts to develop a method of quantifying local blood flow.
Estimating Local Blood Velocity/Flow
Fluid flow in vessels (pipes/tubes) can be obtained from the cross-sectional area of the vessel and the velocity of fluid flowing through that cross-section using the relationships in Eq. 1.2.
= i4rea (cm2) Velocity ()
Q = j> V{A) dA Equation 1.2 Flow from Area and Velocity
Vessel cross-sectional area and centerline direction can be obtained from the physical structure of the vessel (shape of the pipe) which can be obtained from the patient using 3D ultrasound. Black and White (BW) ultrasound data can provide structural information from tissue density differences. Color Doppler ultrasound can provide vessel
Volumetric Flow
cm~
9


structure if the image is acquired with no color blooming outside the vessel lumen since the outer surface of the color volume would represent the inner surface of the vessel as shown in Fig. 1.7.
Figure 1.7 Typical Doppler Color Ultrasound Image of Hepatic Veins
The velocity of the blood in the vessels can be obtained from the Color Doppler velocities if they are Angle Corrected for the vessel direction relative to the transducer beam direction.
10


Challenges/Risks
The customary procedure for ensuring that Color Doppler and Power Doppler data volumes represent the interior volume of a vessel involves: 1) adjusting the ultrasound power and 2D BW gain to levels that show vessel lumens that are visible enough to be seen but do not vary visibly in position or size with excess signal power or gain; 2) increasing the Color or Power Doppler until the color just fills the lumen interior. The accuracy of these adjustments are inherently variable between ultrasound carts and operator judgment and skill. Success also depends on sufficient echogenicity from the blood to reflect enough sound energy for the rendering of a luminary surface.
Hypothesis: Localized patient specific blood flow can be accurately estimated from hepatic 3D Color Doppler data volumes
11


Background Studies
Background studies were performed in the areas of hepatic physiology, hepatic blood flow, blood properties, hepatic ultrasound and vessel reconstruction in order to develop aims supporting the hypothesis.
Hepatic Physiology
The location of the liver in the circulatory system is shown in Fig. 18.
S^via S. Mader, Inquiry into Life, 8th edition. Copyright 1997 The McGraw-Hill Companies, Inc. All rights reserved.
Cardiovascular
System
carotid artery (also subclavian artery to arms)
pulmonary vein aorta
heart
mesenteric arteries
digestive tract
renal artery
iliac artery
trunk and legs
head and arms
jugular vein (also subclavian vein from arms)
pulmonary artery
superior vena cava
inferior vena cava
hepatic vein
liver
hepatic portal vein
renal vein iliac vein
Figure 1.8 Location of the Liver in the Circulatory System
12


The liver takes in low pressure non-pulsatile venous blood from the Gastrointestinal tract as shown in Fig. 19 and passes it back to the heart after rebalancing the blood chemistry and storing energy for use during sleep.
Hepatic Portal Vein Tributaries
Copyright 1997-8 Novarts AJI nghts 'eserveo
Figure 1.9 GI Vasculature Feeding the Portal Vein
13


The anatomy of the liver is shown in Fig. 1.10 where the incoming venous blood enters the liver through the portal vein which branches down to the capillary level to pass
the blood through the liver cells (Hepatocytes).
Internal Anatomy of Liver
Figure 1.10 Vascular Anatomy of the Liver
14


The hepatic veins of the liver collect the blood processed by the hepatocytes and return the blood to the heart through the inferior Vena Cava. The Hepatic Artery provides oxygenated blood to the liver to supplement the de-oxygenated blood coming from the GI tract. The liver also creates bile for storage in the Gall Bladder which is used to emulsify fats in the GI tract.
Figure 1.11 is an X-ray image of the liver showing the internal vascular structure.
https://research,brown.edu/images/pimages/1232327670,jpg
Figure 1.11 2D Image of Portal Vein and Hepatic Artery
15


The surgical regions of the liver are shown in Fig. 1.12. as divided into eight volumes that follow the major branches of the vasculature and allow volumes containing malignancies to be resected without damaging the organ. The liver can completely regrow from 25% of its original structure and mass.
http.//www.cDmc.org/advanced/liver/patients/topics liver-cancer-profile.html
Figure 1.12 Sections of the Liver
16


The primary cells of the liver (Hepatocytes) are organized into repeating structures of six sided Tobules as shown in Figs. 1.13-14 where the organ level
plumbing of the portal vein, hepatic artery, bile duct, and hepatic veins are shown
scaling down to the cellular level.
Centrilobular
Hepatic
vein
vein
Hepatic
vein
00 0
Hepatic
artery
Lobule
Biie Hepatic Hepatic duct artery portal
vein
https ://courses, stu, qmul,ac,uk/SMD/Kb/microanatomv/d/alimentarv/index,htm
Figure 1.13 Hepatic Lobule Structure
17


Lobule
Each lobule contains specialised cells that perform all of the liver's chemical functions
Branch of portal vein
Nutrient-filled blood from the small intestine enters each lobule through sin small veins
Branch of hepatic artery
Owygenated blood from the heart reaches each lobule through sin small arteries
Central vein
Deouygenated and processed blood from the lobule collects in this vein before returning to the heart
Sinusoids
Blood is processed by hepatocytes as it passes along these channels to the centre of the lobule
Branch of bile duct
Bile made by liver cells flows into tiny channels that converge to form the common bile duct
http://www.aviva.co.uk/library/images/med_encyclopedia/cflig5761ivgal_005.gif
Figure 1.14 Hepatic Lobules Flows
The portal vein, hepatic artery and bile ducts follow the edges of the lobules and are referred to as the Triad. The portal vein and hepatic artery blood flow from the triad through sinusoids lined with hepatocytes to reach the central veins (acini) that are attached to the hepatic vein as shown in Figs. 1.15-16. Bile created by the hepatocytes is removed though the bile canaliculi.
18


http://www.ouhsc.edu/histologv/Glass%20slides/88 01 .jpg
Slide 88 Liver
£fcntratvein,
PortaU'r lad v \ with surrounding connective tissue
Figure 1.15 Lobule Cross-section
Livrr
I-------------------------------------
SiiriisrticJ
Scpt.il KupFfai cell Hup-itm yl-
Central ve.n
Bile
ranakculuy
Sinusoidal lumen Kuoffer
Space ol L>iyyc
.all
Dendritic:
wiMi ieliestlae La1t
Hepatic anery t*i.i ii h
vein Li am h
www.nature.com
Figure 1.16 Detail of the Blood Flow through the Hepatic Sinusoids


The hepatocytes nearest the portal vein inlet from the triad are most likely to be damaged by toxins (accidental acetaminophen overdose) or de-oxygenation. There are other cells in the liver with specialized functions. Stellate cells (Ito) store fat and Vitamin A and produce collagen. Kupffer cells are specialized macrophages which remove bacteria and debris from the blood. Dendritic cells are immune cells that process antigens.
Hepatocytes are 30-40 um wide and line the walls of the sinusoid. A schematic of the cell is shown in Fig. 1.17.
20


canaliculus occludens apparatus of Diase reticulum reticulum Glycogen
Endothelium
Glycogen
Golgi i Zonula
apparatus I occludent;
Smooth
endoplasmic
Bough
endoplasmic
reticulum
reticulum
Mitochondria canaliculus
Lumen o! sinusoid
ELSEVIER, INC. ELSEVIERIMACES.COM
Figure 1.17 Hepatocyte Cellular Structure
In summary, the liver is situated in a major blood collection path to enable the
cleansing of a continuous fraction of venous blood and the storage of energy. The large size of the liver facilitates a low blood flow and low pressure drop which ensures that the hepatocytes have enough time to interact with the passing blood.
21


Hepatic Flow Rates
To gain a better understanding of normal hepatic flow rates a literature search was
performed and the findings are compiled in Table 1.1.
Table 1.1 Normal Portal Vein Size, Flow, and Pressure Characteristics
Flow (ml/min) Pressure (mmHg) Diameter (mm) Area (cm2) Mean velocity (cm/s) # Age M/F Ref
min mean max min max min mean max mean min mean max
899 0.99 15.3 85 [26]
724.2 12 4.524 14.2 10 21-55 f [27]
1000 1200 6 [28]
9.8 10.5 11.2 15.7 21.8 27.9 19 20-40 m [29]
7.7 8.3 8.9 26.3 32.5 38.6 18 20-40 f [29]
9.6 43 adult m/f [29]
8.9 36 22-75 m/f [29]
9.4 37 20-40 m/f [29]
11.7 30 24-63 m [29] [30]
11 30 41-75 m/f [29] [31]
847 1.27 11.08 16 18-58 m [32]
715 1 12.43 13 18-58 m [32]
878 1.01 14.4 19 21-58 f [32]
613 0.75 15.26 14 21-58 f [32]
5 8.5 12 21 0-10 m/f [33]
7 10 13 20 11-20 m/f [33]
6 11 15 49 21-30 m/f [33]
6 11 15 58 31-40 m/f [33]
11 107 21-40 m/f [33]
13.6 32.15 55 29 23-65 m/f [34]
7 12 [35]
The flows are reported for the portal vein and do not include hepatic arterial flows. The data in the table reflects that diameter and velocity are easily obtained from non-invasive ultrasound measurements using caliper tools in 2D and Pulse Wave Doppler ultrasound. Pressures are much more difficult to obtain as they cannot be directly measured in human subjects and must be inferred by techniques such as the Hepatic Venous Pressure Gradient (HVPG) [36], explaining the sparse data entries in the table.
22


The hepatic pressures reported in Table 1.1 are in relative agreement with the
Venous Pressures suggested by the annotation in Fig. 1.18. at 5-7 mmHg.
Figure 1.18 Operating Inlet Pressure of the Liver
23


The normal hepatic operating ranges are shown together in the diagram of Fig. 119.
Figure 1.19 Diagram of the Hepatic Flows and Pressures [28]
The hepatic veins collect the sum of the portal and hepatic artery flows (1200 + 400 = 1600 ml/min) and as the adult liver volume is roughly 1600 ml, the liver blood is changed about once per minute. As the adult human body contains roughly 5 L of blood (4.57L F/5.32L M) [37], the liver could theoretically process the entire body blood volume within 5 minutes, however, it takes much longer since the liver is only one of many venous returns to the heart and the GI tract is only one of many recipients of aortic blood. There is a small pressure drop from 7 mmHg at the portal inlet to 4 mmHg at the hepatic vein connection to the inferior vena cava [24], The portal flow is non-pulsatile having come from the venous side of the GI tract, whereas the hepatic artery flow is highly pulsatile as it branches from the descending aorta just below the heart. Heat added
24


to a liver tumor during the RF ablation raises the temperature of the surrounding tissue and local blood. Estimating this heat distribution will be treated in follow-on research.
Blood Properties
Blood consists of plasma (55%), red blood cells (44%), and white blood cells by volume (<1%). Hematocrit is a measure of the red blood cell count to the total blood volume and has a value near 45%. As plasma is primarily water (92%) and contains a low content of albumen (7%) [38] [Red Cross] it is considered a Newtonian (ideal) fluid where the shear stress (t) and strain rate (d9/dt) are proportional, or, x/(d0/dt) = 1. As whole blood contains deformable red blood cells and other cells it exhibits viscoelastic properties depending on the vessel size. Viscoelasticity refers to the ability of the 2x7 um dumbbell shaped red blood cells to deform elastically to fit through capillaries, thus changing the blood viscosity as shown in Fig. F20.
25


Figure 1.20 Viscoelastic Properties of Blood at 2 Hz and 22 C
In vessel diameters above 0.5 mm, whole blood behaves as a Newtonian fluid and below that diameter whole blood behaves as a non-Newtonian shear thinning fluid where the shear stress exceeds the strain rate, x/(d0/dt) > 1. The dynamic viscosity of whole blood also varies with temperature and hematocrit [39] and measured values are provided in Table 1.2.
Table 1.2 Blood Properties
Substance Dynamic Density Conditions Ref
Viscosity (p) (kg/ms) kg/m3
Blood 4E-3 1060 37C 45% hematocrit [39]
Blood 1058 [40]
Water IE-3 [38] [39]
Water 992.2 40C [various]
26


The thermal characteristics of blood are shown in Table 1.3.
Table 1.3 Thermal Characteristics of Blood and Liver
Substance/Tissue Thermal Conductivity (W/mK) H20 (%)
Aortic .476 54
Whole .492 54
Plasma .570 93
Pure Water .627 too
Liver .469/. 564 77
Tumor peripheral .511
Tumor core .561
Hepatic Flow Dynamics Hepatic Resistance
As illustrated in the radiograph of Fig. 1.11, the liver distributes portal vein inflow to the hepatic sinusoids at the capillary/cellular level and the hepatic veins re-collect the flows for return to the inferior vena cava. While modeling the resistance of flow at the branching and sinusoidal levels would be very complex, a rough figure of resistance can be obtained by treating the liver as a control/storage volume and using a reduced form of the Windkessel equation (Eq. 1.3) with the following assumptions; portal flow is 75% of the total in-flow, hepatic flow is 25% of the inflow, inflow equals outflow after the liver is full of blood, steady flow, no change in liver volume, portal pressure of 7 mmHg (933.25 Pa), outlet pressure of 4 mmHg (533.29 Pa), and from Table 1.1 an average flow rate of 800 ml/min (13E-6 m3/s).
27


Windkessel (AP/dt = 0)
Q =
A P dAP
T + k~
Rh
Ppv RflV
933.25 533.29 13-6/0.75
Pa
= 23.1E6-
m-
= 2.9E6 Wood Units
Equation 1.3 Hepatic Resistance by Windkessel
In clinical practice, a Resistive Index (RI) is computed from the systolic and diastolic velocities of the hepatic artery using Eq. 1.4.
DT ^Sys Vj)jas
RI" vSys
Equation 1.4 Resistive Index for Hepatic Artery
Hepatic Portal Vein Flow Profile
Modeling the portal vein as a simple smooth pipe flow, the Reynolds number provided by Eq. 1.5 is an indicator of laminar (Re < 2300 ) vs turbulent (Re > 6000) flow [41], Laminar flows take on a parabolic velocity profile across the vessel cross-section
28


downstream from curves and stenosis (Poiseuille), where turbulent flows exhibit a uniform velocity profile (Plug).
Using the density of blood of 1060 kg/m3 and a dynamic viscosity of 4E-3 kg/ms from Table 1.2, and, an average portal vein diameter of 10 mm (.01 m) and velocity of 15 cm/s (0.15 m/s) from Table 1.1.
p v D Re = -----
p
_ 1060x0.15x0.01
~~ 4xl0-3
= 397.5 (< 2300)
Equation 1.5 Portal Vein Reynolds Number
The result from Eq. 1.5 is well below the approximate threshold for laminar flow suggesting that the portal vein flow is laminar.
Pathological Flow
The principle inhibitor to flow through the liver would be a diagnosis of cirrhosis, a fibrosis scarring around hepatocytes which have been injured or killed by disease (Hepatitis B) or toxins (acetaminophen) that increases the resistance to blood flow through the damaged sinusoids as shown in Fig. 1.21.
29


Figure 1.21 Cirrhosis
This increased resistance to flow increases the pressure drop across the liver and can lead to a condition of Portal Hypertension which is clinically indicated by a portal vein pressure greater than 12 mmHg [31] or a portal velocity less than 16 cm/s [28], The velocity indicator is not conclusive alone as it overlaps the normal ranges listed in the studies of Table 1.1.
Hepatic Ultrasound
As ultrasound is a low cost imaging modality and a 3D system was available for this project, the characteristics of liver and surrounding tissues was compiled in Table 1.4.
30


Table 1.4 Hepatic Ultrasound Characteristics
Tissue Density (kg/m3) c (mm/uS) Attenuation (dB/cm) @ 1MHz Impedance (z) (rayls) Ref
Liver 1061 1.555 .4 .7 1.65E6 [40]
Liver 1060 1.566 1.66E6 [39]
Blood 1058 1.560 .18 1.65E6 [40]
Blood 1060 1.566 1.66E6 [39]
Fat 924 1.450 .5 1.8 1.34E6 [40]
Fat 920 1.446 1.33E6 [39]
Muscle 1068 1.600 .2 .6 1.71E6 [40]
Muscle 1070 1.542- 1.65- [39]
1.626 1.74E6
As Color Doppler velocities are relative to the origin and direction of the ultrasound beam, it is necessary to perform an angle correction between the direction of the vessel centerline and the direction of insonation as depicted in Fig. 122.
Figure 1.22 Geometry of Doppler
31


In current practice, the angle correction procedure is performed manually on the screen of the ultrasound cart by the clinician for each location of interest along the vessel as shown in Fig. 1.23 by placement of a sample volume box in the center of the vessel and rotating a cursor in-line with the vessel flow.
Figure 1.23 Example of Manual Angle Correction
Vascular Segmentation
Hepatic Vascular Structure
Many images of hepatic vascular structure were viewed to gain insights into the challenges of vascular reconstruction. The best physiological example of the finely arborized hepatic branching is shown in Fig. 1.24.
32


Figure 1.24 Hepatic Branching (Google Images)
Segmentation Algorithms
The search terms liver, vessel, 3D, ultrasound, reconstruction, snake, extraction, segmentation, and tracking were used to find relevant literature. The most recent published review of vascular reconstruction techniques uncovered to date is Kirbas [42] in 2004 which is an exhaustive work that categorizes reconstruction methods by approach. Twenty eight other more recent papers were reviewed and sixteen were found to be relevant to the mission of autonomous vascular reconstruction from 3D ultrasound as there are many other segmentation algorithms that require the user to manually locate seeds [43] [44] [45] and otherwise manually direct the progress of the segmentation process. Segmentation approaches that rely on pre-trained data or use anatomical libraries were reviewed to uncover ideas for fully automated approaches.
33


Of all papers reviewed, only two and were specific to 3D ultrasound [46] [47], providing an apparent opportunity to advance vascular discovery with this lower cost modality.
Most of the reviewed articles employed processes that benefit from venous contrast agents in conjunction with MRI and CT modalities. Only one paper treated zero contrast applications [48], As contrast agents are not FDA approved in the United States for ultrasound, the extraction of vasculature from Ultrasound images is more challenging [46] [47] [48],
As with real hepatic anatomical structures, vessels captured in 3D images do not spatially overlap each other, eliminating the need to process for overlapping structures as in 2D processing [45], however, wave front propagation has been proposed as a means of detecting the centers of tubular structures which may be overlapping in 2D or close to each other in 3D [45],
References discuss the issues with scalability of some approaches and offer multi-scale alternatives [45, 49-52], Scaling problems occur when a filter must be optimized for the size of a structure and gives less conclusive results with larger or smaller versions of the structure.
All papers indicate that the first processing step is to reduce the random noise in the image since the accuracy and decision making of all tracking algorithms will be compromised by voxels which are not representative of the tissue at that location. This noise specifically allows region-growing algorithms to bleed beyond vessel walls.
Noise reduction techniques included Mean-Gaussian [44] and Diffusion filtering[53],
34


Intensity thresholds were used by some processes to remove non-vessel voxel noise from image volumes [49]
Ultrasound data contains speckle noise [47] [27], a natural artifact of a coherent acoustic wave reflecting from objects and impedance discontinuities. This noise can be either treated as a noise to be eliminated through filtering, or, may be processed to provide additional information such as the density and number of random scatters based on the statistics of the received envelope [46], This latter treatment of speckle is useful in the discrimination of breast cancer.
Discovery of hepatic vessel interiors can be performed by growing a manually or automatically placed seed volume into a larger volume which is constrained at the vessel boundary by voxel intensity changes or other rules based on force, entropy, energy or advection [43] [51],
Many previous investigators use a Hessian matrix as a filter to determine how similar an arbitrarily oriented structure is to a vessel [48] [45] [49] [52] [50] [54] [55], The Hessian matrix is trained to the characteristic values of a tube structure and the eigenvalues of the structure are used to provide an indication of vessel-ness. One process implements multiple eigenvalue sets to accommodate the identification of both tube-like and junction-like structure [54], Hessian filters deal well with irregularly shaped voxel dimensions (anisotropy).
When a 3D image is taken as slices, the vessel cross-sections appear as individual circles and ellipses. The centers in each slice are determined by graphical or mass centroid or core methods [56], Vessel tracking may be accomplished by treating the
35


movement of the vessel center as target tracking. This class of problem can be treated with Kalman filtering, a process developed for predicting the trajectories of space objects and military targets [45] [47], The Kalman filter process uses the present and previous states (center locations in each slice) of each circular object to predict the center location in the next slice. As each new slice is processed and the new centers can be known, the weighting factors of the filter are adjusted to more accurately predict the next center location. The Kalman filter is very accurate for gradual transitions and less accurate in sharp or reversing bends, a situation common in hepatic vasculature [47],
A number of papers present reconstruction by use of the intensity based voxel by voxel wise area and volume search. B-splines have been evaluated for establishing hepatic vessel centerlines as they establish a minimum curvature fit [53], A geometrical moment method using a cylindrical shape as a vessel reference for reconstruction has been evaluated [56],
The tabulation of the approaches provided in Table 1.5 provides a summary of approaches to vascular reconstruction, however the documentation of success with ultrasound as a noisy and Tow contrast imaging modality was sparse.
36


Table 1.5 Summary of Vascular Segmentation Approaches
Ref Modality User Input Noise Reduction Approach Vessel Segmentation / Detection Centerline Construction Vessel Surface Rendering
[43] CT Seed and Contour Points Yes, not specified Region Growing N/A Skeletonization Surface of Region
[44] CT Editing Mean- Gaussian Graph-cuts Centerline extraction Tree separation Medialness filter Radius from Medialness filter
[45] 2D (3D) Xray No Not discussed Minimum cost Hessian Wavefront 3D not treated
[48] CT Gaussian Median No contrast Hessian Thresholding Hessian
[471 US Seed Kalman Median filter
[27] US -Peak/Valley -Hilbert Curve -Cubic Spline -Windows adaptive threshold Otsu Core area Morphological filter
[53] MRA 3D Diffusion Filter Voxel search 2ntl Order B-spline
[49] CT Threshold Hessian -Hough transform -Morphology
[51] CT Edits Threshold Region Growing Hessian Hessian
[521 MRA Hessian Hessian
[50] MR Hessian Difference Hessian
[54] CT Correlation Multiple Hessian Morphology Fuzzy shape
[551 MRA Gaussian Multiscale Hessian
[56] MR/CT Seeds Geometrical Moment Boykovs Graph-cut Min cut Max flow
Preliminary Data
Segmentation Applications
Existing techniques for vessel segmentation and surfacing were explored using available software applications such as Geomagic and MeshLab as documented in the
37


Portal Vein Flow Phantom fabrication (Chapter IV) and the following discussion of an ITK-SNAP advection approach.
Advection is a method of segmentation where a boundary between tissues or fluids is defined by the user as a specific change or rate of change in voxel intensity. A balloon-like 3D manifold surface is then inflated/expanded from a user selected seed point until the expanding surface reaches a boundary that meets the threshold or gradient rule and then stops at that voxel location. When all of the voxels of the expanding surface reach a boundary, the final expanded surface represents the geometry of the boundary surface and has an identical volume and surface area to the boundary surface.
The Advection capabilities of ITK-SNAP (www.itksnap.org) were evaluated on several arterial and venous phase CT image volumes taken from liver donor studies.
Liver donor studies were used since the success of a transplant relies on good structural views of the donated organ in surgical planning. CT imagery was used for the evaluation since it is a much less noisy and higher resolution imaging modality than ultrasound, providing a benchmark of how well advection would perform on a good image. The reasoning was that if advection yielded marginal segmentation on CT imagery, it would not be acceptable for ultrasound.
As a contrast agent is used in the liver donor image acquisition, the contrast in the arterial phase provides a distinct vessel wall as shown in Fig. 11.25.
38


Figure 1.25 Arterial Phase CT
An advection segmentation of the descending Aorta and hepatic artery during the Arterial phase shown in Fig. 11.26, where the hepatic and splenic artery surfaces are well defined and reasonably smooth.
Figure 1.26 Arterial Phase Hepatic Artery
In the venous phase, the contrast agent has been diluted throughout the blood
stream and filtered by the liver hepatocytes, resulting in lower image contrast and as shown in Fig. 11.27.
39


Figure 1.27 Venous Phase CT
The advection of this low contrast image volume provided a much less accurate and unusable vessel wall definition (Fig 11.28) due to the advection surface leaking into noise artifacts along the vessel wall.
Figure 1.28 Portal Vein
40


The advection process is unable to distinguish a legitimate voxel intensity change at a physical boundary from the same intensity changes between voxels due to noise.
Even using an established and well developed tool like ITK-SNAP on a fast laptop, the advection process took tens of minutes to complete which is not in line with the required TTR. The advection approach was not pursued further due since ultrasound imagery is more noisy than CT and the clinical time requirement was not met.
Clinical Requirements
Software Application
A graphics based software application was the natural vehicle for the development of the algorithms required to estimate flow from Color Doppler and to capture their value in a clinically useful manner. As such, there were clinician requested features directing usability and response time of the software application. The first directive was that the clinical application implement a true 3D display that would allow vascular images to be viewed from any angle or perspective. The second directive was that all functions other than the capture of a Region of Interest (ROI) around a tumor would operate autonomously without user direction or intervention. The third directive was that the clinical application adhere to a Time-to-Response (TTR) of 10 seconds to manage the cost and timeliness of the RF ablation procedure. The TTR directive reflects that: clinician time is costly and a precious commodity; Operating Room rates can reach
41


thousands of dollars per hour; and that the procedure is performed on a conscious patient who would like to leave as soon as possible.
Review of Existing Tools
Prior to the development of custom software, existing tools were reviewed for compliance with the research and clinical goals. The review included Osirix (Mac), ImageJ, Irfan View, 3D Sheer, Syngo FastView, Onis, ORS, Qlab (Philips), ANSYS, SolidWorks, Meshlab, and Geomagic. Ultimately, the Philips QLab software application was the only available application capable of opening the 3D files from the available iU22 ultrasound cart. While QLab did provide the desired 3D viewing and image manipulation, and manual and 2D automated segmentation, the application did not offer vessel center-lining, automated angle correction or flow estimations.
Software Architecture MATLAB
MATLAB was selected for initial algorithm development and prototyping. MATLAB was not used in the development and deployment of the clinical algorithms and application due to the TTR requirement and Intellectual property concerns. MATLAB is based on JAVA, an interpreted language that requires each line of MATLAB code to be broken down into JAVA code and then run on the JAVA Virtual Machine (JVM) every time a program is run. MATLAB documentation suggests that time critical applications be developed in a compiled language such as C where the
42


conversion to machine executable code happens only once. While MATLAB offers a C cross-compiler, the underlying library algorithms are not necessarily converted into the most efficient constructs of the C language, and, the code-compiled library functions carry an IP burden. As a collection of pre-developed libraries, MATLAB retains the rights to them through an expensive commercial license, a cost and burden that the project was not willing to bear. While MATLAB does allow applications to be deployed using their client at no cost, it was deemed not feasible to add this client software to the laptops used in the hospital IT environment.
The C Programming Language
Other than being license free and widely supported, the most attractive feature of C is that it allows the direct and efficient manipulation of data in memory through the use of pointers and array subscripts, critical to fast processing of large images in real time clinical settings. Memory access in object oriented languages such as C++ or JAVA is performed through special class functions, which slow the process. For nonmemory functions, C++ has features which are very useful for Graphical User Interface (GUI) implementations and general control of applications and it integrates well with native C code.
For the reasons stated above, the clinical application was implemented using the C programing language. To keep the clinical application development process manageable, the process was broken into two steps: 1) implement the vascular
43


reconstruction algorithms and basic graphics functionality in C and OpenGL; 2) migrate the application to the C++ based Qt user interface.
Since the C programming language does not define or include a library of functions for the management of graphical data or user interfaces, the OpenGL (www.opengl.org) graphics library was selected. OpenGL is a free to use and distribute library when dynamically linked by the application (static linking requires a distribution license), it is supported by all major operating systems, and is deployed on thousands of graphics systems and mobile devices.
To provide the GUI and mouse based user controls and measurement reporting capability expected of a clinical application, the C/OpenGL code base was imported to a user interface development environment called Qt. Qt was developed by Nokia for deploying interactive graphical applications across many stationary and mobile devices (laptops, tablets, cell phones) and operating systems (Windows, Linux, Android,
MacOS). Qt was chosen to allow the clinical application to be deployed to laptops, tablets and phones with relative ease and consistency. Google Earth is an example of a successful cross-platform Qt application. There are no Qt licensing fees if the library functions are linked dynamically (i.e. DLLs), however, a distribution license is required if a static library is to be embedded within the application.
TTR Compliant Architecture
As Operating Room and clinician overhead costs are high, the algorithms and applications were designed to minimize processing and calculation times. The
44


application architecture is based on fixed size pre-allocated static integer arrays for one time storage of the raw image data. This eliminates the time and need to re-scan the image data from files and the time required to dynamically allocate and de-allocate the very large arrays which are up to 1720 x 1024 x 180 x 3 Bytes (1TB) for the Visual Human Images and 512 x 512 x 512 x 3 Bytes (500 MB) for 3D ultrasound data. A side benefit of static arrays is that they can be successfully implemented under any Operating System (OS) (a clinical deployment goal). Many Operating Systems do not have robust allocation/de-allocation functions, which can result in memory leakage and system crashes that make the application appear unstable and unusable. The benefit of using a static integer array is that it is smaller than using a floating point data type. As the volume of vessel structure is a very small percentage of the total image volume, the vessel cross-sections and properties are also stored as objects in a list (a single dimension linear array) as they are discovered. With raw data available in the static 3D array, functions that involve relative and absolute indexing to voxels of interest can be performed very quickly with integer math, and, vessel objects, features and parameters can be quickly picked from a list. This combined approach enables very quick calculations and the best possible user response times.
45


Development Environment
To facilitate speed of development, a Dell Precision M4500 laptop PC was used with the following configuration: Operating System Windows 7 64 bit, CPU Quad Intel 17 2.4Ghz, RAM 8GB, Graphics NVIDIA Quadro FX 1800M; Programming Environment MATLAB, Qt, DevC++; CAD Applications MeshLab, Geomagic, SolidWorks, ANSYS.
46


CHAPTER II
RESEARCH
Specific Aims
The development of autonomous flow estimation from 3D Doppler ultrasound was broken into steps to guide the development and verification of individual algorithms.
Aim 1 Fabricate and Model a Portal Vein Flow Phantom
In order to validate hepatic flow estimates it was necessary to create anatomical and geometric flow phantoms, and perform flow modeling.
Aim 2 Obtain an ROIfrom a 3D Image Volume from Clinician Input
To begin the flow estimation process, it was necessary to acquire 3D image data from an ultrasound cart and provide the clinician with a graphical means of specifying what Region-of-Interest (ROI) of the 3D volume contained the tumor of interest. Estimating the flow within a sub-volume of the entire 3D image was deemed a more achievable and more TTR compliant task.
Aim 3 Autonomously Segment Vasculature and Compute Centerline
With an ROI selected, it was necessary to find and identify all of the unique vessel segments within the ROI accommodating vessels that may be very close together
47


and be in proximity to spatial noise artifacts. It was also necessary to find each vessel segments centerline to facilitate diameter determination and angle correction of velocity.
Aim 4 Autonomously Compute Segmented Vessel Diameters and Cross-sectional Areas
The next step in flow estimation involved finding the vessel diameter from its cross-section (which is normal to the centerline direction) and computing the area of that cross-section.
Aim 5 Autonomously Compute Angle Corrected Velocities for Segmented Vessels Because Color Doppler measures velocity relative to the beam direction, it was necessary to correct the Color Doppler velocity data in the direction of the vessel at the point of intersection on the vessel centerline to obtain the velocity in the direction of the vessel. This required a determination of the transducer origin for angle correction with the centerline estimated in Aim 3.
Aim 6 Autonomously Compute Volumetric Flow for Segmented Vessels
Estimated flow was obtained by multiplying the angle corrected velocity and cross-sectional area at the points of interest along the centerline
Each algorithm developed in the Aims above was evaluated using flow phantoms of known physical geometry and calibrated flow rates.
48


Materials and Methods
Aim 1 Fabricate and Model a Portal Vein Flow Phantom
Portal Vein Phantom
As a study of hepatic vascular flow, industrial casting materials were researched and imaged with ultrasound, MRI and CT, and then used to fabricate flow phantoms.
Materials Study
A set of usability and performance criteria were established for a proposed phantom design capable of supporting liquid flow during imaging. A literature search was conducted to identify the materials and methods previously used in phantom fabrication. A database of human tissue and casting material properties was compiled to facilitate the selection of appropriate materials for testing. Several industrial casting materials were selected, procured, and used to fabricate test samples that were imaged with Ultrasound, MRI and CT. The full text of the published phantom materials study [57] is provided in Chapter III.
Fabrication
A portal vein structure was extracted from the Visual Human Male data set for use as a pattern in the flow phantom construction. A three-dimensional (3D) portal vein pattern was created from the Visual Human database. The portal vein pattern was used to fabricate two flow phantoms by different methods with identical interior surface geometry using Computer Aided Design (CAD) software tools and rapid prototyping techniques. One portal flow phantom was fabricated within a solid block of clear P-4
49


silicone for use on a table with Ultrasound or within medical imaging systems such as MRI, CT, PET or SPECT. The other portal flow phantom was fabricated as a thin walled tubular dipped latex structure for use in water tanks with Ultrasound imaging. Both phantoms were evaluated for usability and durability. The full text of the published phantom fabrication [58] is provided in Chapter IV .
Image Testing
The image quality of the perfused solid silicone flow phantom was tested on a lab bench with the 3D transducer coupled to the phantom body with acoustic gel as shown in Fig II I.
Figure II. 1 Solid Silicone Body Portal Phantom and Test Setup
50


The image quality of the perfused thin walled latex phantom was tested in a water
tank with the transducer and phantom immersed as shown in Fig II.2.
Figure II.2 Thin Walled Latex Portal Phantom and Test Setup
Portal Phantom Flow Testing
The branch outflows of the thin walled latex portal phantom were measured using a mass/time method over a wide range of flow rates and pressures taken from Table 1.1 and the test setup shown in Figs II.2-3. The adjustable height (0- 32) gravity tank used to adjust head pressure is shown on the left of Fig. II.3 and the individual branch outflow tubes are shown on the right.
51


Figure II.3 Thin Walled Latex Portal Phantom Test Setup
The water outflow of each phantom branch was collected over a fixed period of time and was then weighed to provide mass flow rates in kg/s, providing a mean mass flow rate.
CFD Simulation
A CFD flow model and simulation was created using ANSYS CFX software. The portal vein flow phantom mesh documented in Chapter IV was imported as a pattern dependent mesh into the ANSYS ICEM mesh editor so that all surface inaccuracies in the portal mesh would influence the simulation results. The inlet boundary conditions were set to flow rates and pressures taken from the portal phantom testing as discussed above and shown in Figs. II.2-3. The branch outlet boundary conditions were set to a constant
52


backpressure representing the 9 (22.9 cm) of head height in the portal vein branch outflows. Water was selected as the fluid and the default ANSYS properties were used.
Portal Vein Phantom and CFD Comparison
The results from the measured phantom flows and the CFD simulation were converted to volume flow rates and plotted together for comparison. The mean Mass flow rates from the branch measurements were converted to a mean Volume flow rate using Eq. II. 1 and the density of water at 0.001 kg/cm3.
Pulse Wave Doppler velocity measurements were not included in the comparison due to the lack of reliable physical registration between the CFD simulation and PW Doppler measurement points along the phantom vessels.
Verification Phantom Setup
The clinical version of the software application (VAS) and developed algorithms were evaluated using simple flow phantoms comprising tubing of known diameter at varied flow rates and transducer beam sweeps. Water was used as a test fluid and dry corn starch was dissolved to increase the echogenicity of the solution.
k Cf
Water Density ()
Mean Volume Flow
Equation II. 1 Mean Mass to Volume Flow Conversion
53


Geometrical Tubing Phantoms were fabricated using latex tubing of varied sizes
to enable testing of individual and grouped tubes as shown in Fig. II.4.
Figure II.4 Thin Tube Phantoms
The parallel phantom is show in the left image of Fig. II.4 where the 3/8
(0.925 cm), l/4(0.635 cm), and 1/8(0.317 cm) latex tubing segments were connected to the inlet and outlet manifolds with independent valves controlling the flow through each tube. The parallel phantom was used for practicing the capture of non-aliased Color Doppler flows utilizing the full velocity scale of the ultrasound cart.
The series phantom is shown in the middle image of Fig. II.4 where the 3/8, 1/4 and 1/8 latex tubing segments were connected end to end with the 3/8 accepting the inlet flow and the 1/8 tube carrying the outlet flow, with all flow controlled by an outlet valve. The series phantom was used to investigate the color range of the Color Doppler velocity scale.
54


The net flow phantom is shown in the right image of Fig. II.4 and used a 3/8 latex tubing with a 1/16 wall in a serpentine layout to allow the same flow to transit the image volume in both directions. The net flow phantom was used to test the clinical applications ability to measure vessel diameter, vessel area, angle corrected velocity, inlet and outlet flow rates and, and zero net flow through the ROI.
Net Flow Phantom Calibration
Two methods were used for measuring steady flow through the net flow phantom and the results from several image samples were compared at several flow rates.
Mass/Time
The mass of water exiting the net flow phantom over 2 a minute interval was measured on a scale to provide a mean Mass flow rate. The time interval was measured with a stopwatch. The mean Mass flow rate was converted to a Volume flow rate using Eq. II I.
Angle Corrected Velocity
Using the ultrasound carts Pulse Wave (PW) Doppler Sample Volume and Angle Correction tool, the Angle Corrected Velocity of the fluid in the phantom tube was measured at several locations along the flow in cm/s. The Volume flow rate was obtained by using Eq. II.2 using the phantom tubing diameter of 3/8 (0.925 cm).
55


/cc\ (cm\ (Diameter (cm)\2
Volume Flow yJ = Velocity yJ *n I----------------------I
Equation II.2 Velocity to Volume Flow Conversion
The measurements from the ultrasound carts were: 1) Time Averaged Mean Velocity (TAMV) obtained by averaging all of the highly variable velocity data; and 2) Time Averaged Peak Velocity (TAPV) obtained by averaging the maximum velocities (velocity peaks). For laminar and parabolic profile flow (Poiseuille), the peak (centerline) velocity is expected to be 2x the mean value. For turbulent and flat profile flow (Plug), the peak and mean velocities are expected to be the same.
Flow Estimation Verification
To expose issues with the velocity estimation methods, five randomly selected ROIs were taken from images of the tubing in the net flow phantom which were acquired under different flow conditions. The images of the net flow phantom were taken with the transducer centerline at an angle of 45 to the net flow phantom tubing with the beam scanning +/- 30-40 from the transducer centerline. To utilize the maximum dynamic range for each of the samples, the velocity scale and flow rates were adjusted such that maximum scale value was slightly higher than the highest Doppler velocity detected in the image, with a slight margin to prevent aliasing artifacts. The velocity ranges were selected to test the performance of the autonomous algorithms, which may create laminar and/or turbulent flow conditions. The Reynolds numbers of the centerline and cross-sectional mean velocities were both computed to help determine the nature of the flow.
56


The following parameters were measured and recorded:
1) Maximum Velocity Scale value from the ultrasound cart color/velocity display
2) The Average of 3-5 manually angle corrected Time Averaged Mean Velocity and Time Averaged Peak Velocity taken from sample volumes along the net flow phantom tube
3) The Cross-sectional Mean of the Angle Corrected Inlet/Outlet Velocities taken from all of the voxels captured in the cross-sections at the top and bottom of the ROI
4) The Centerline Angle Corrected Inlet/Outlet Velocities taken from the top and bottom centerline voxels within the ROI
Aim 2 Obtain an ROIfrom a 3D Image Volume from Clinician Input
Obtaining 3D Image Data
Access to a Philips (Bothell, WA) iU22 Ultrasound system (SN 02L6HO, Rev. 6.0.0.X) and a V6-2 curvilinear 3D transducer was obtained for the capture of 3D ultrasound images. Philips provided a precompiled (.p) MATLAB library function that allowed access to their proprietary 3D file format.
Displaying 3D Image Data and Capturing an ROI Algorithm Development Platform (ADR)
To facilitate the examination of raw data from the ultrasound cart and development of prototype algorithms, a custom GUI based viewer named the Algorithm Development Platform (ADP) was developed in MATLAB. The ADP Graphical user interface was designed to allow: viewing image data in two viewing panes; viewing data from each of the lateral, depth, and elevation views; data to be viewed in slices
57


synchronized between the BW and Color Doppler data; the development of vessel finding, segmentation, and ROI selection functions. The ADP GUI is shown in Figs. II.5.
Figure II.5 BW and Doppler Slices (BW Mode)
The default mode of the ADP is View mode set by the View/Build button which allows the developer to view the raw image data or data created by the functions developed for vascular reconstruction as described below. The ADP was designed to allow 2D slice by slice viewing of 3D data in all axes (Depth, Lateral, Elevation) of the dataset both Forward and Reverse order using the DF, DR, LF, LR, EF, ER buttons. All operations were started with the Run button, paused with the PSE button, and stopped completely with the Stop button. The rate of slice display was controlled by the Delay (DL) button and the slice count in the image was displayed in a text box. The Slice (SLC) button allowed stepping through the image volume on a slice by slice basis. The
58


Test (TST) button forced the ADP to use pre-configured test image volumes to assist with image registration and orientation. The color mapping of the image data is selected with the COL button and the color gamut was displayed as a bar on the right margin of the GUI. As MATLAB (Windows OS) applies a color map on a global (not window) basis, both panes are forced to the same color map. Images mapped to a linear BW map are shown in Fig. II.5 and images mapped to a linear false color map are shown in Fig. II.6. False color mapping has no meaning for BW images.
Figure II.6 Geometry of Doppler (False Color Mode)
The ADP was designed with two viewing panes with independent and identical controls such that a BW ultrasound image slice could be displayed in one pane and the identical slices of the Color Doppler data could be shown together in the same pane as shown in Fig. II.6, or, the red Doppler data (flow towards the transducer) could be
59


displayed in one pane and the blue Doppler data (flow away from the transducer) could be displayed in the other pane as shown in Fig. II.7.
Figure II.7 Red and Blue Doppler (False Color)
As utility functions, the HST button plots the histograms of color/velocity values within the selected image data as shown in Fig. II. 8.
60


Figure II.8 Histograms of Red and Blue Doppler
The CMP button displays the active color map shown in Fig II.9.
Figure II.9 Linear Color Mapping of BW and Doppler
61


The ROI button creates orthogonal side view and a movable selection box to
allow selection of a 3D ROI.
The ADP was implemented in 2500 single spaced lines of original code.
Vascular Analysis System (I AS) (C/C++/OpenGL/Ot) [UltrasoundData] The VAS user interface is shown in Fig. II. 10.

Figure 11.10 VAS Clinical Application Graphical User Interface
Leveraging OpenGL capabilities, Orthogonal, Perspective and Navigational viewing modes were implemented. The Orthographic and Perspective modes allow the clinician to rotate, spin and translate the image around a fixed focal point at the origin in
62


the center of the vasculature image with control over the eye-to-origin distance. In Orthographic mode, the data is shown with right angles. In Perspective mode the data is shown with a vanishing point as it would be seen in a photograph or painting. Navigation mode allows movement of the point of focus in the scene and allows the clinician to travel towards and away from the focal point as if the viewer was walking through the image volume. Navigation mode is also called fly through mode and allows the clinician to see the inside of vessel surfaces for inspection of stenosis, branching and other vascular pathologies by viewing them form within the structure as a small observer. The Up/Down/Left/Right buttons pan the scene in Perspective mode and shift the focal point in Navigation mode. The Axes button turns the white x, y, z reference axes on and off. The view of the object can be set to an orthogonal view at any time with the Depth/Lateral/Elevation direction from the back or front. The default Orthogonal view can be recovered with the Reset. The Slice button enables the clinician to enter an arbitrary slice, or, increment/decrement by one slice to step through the image. Buttons at the bottom of the display enable toggling independent display of the Black and White, Red Doppler, Blue Doppler data.
Since OpenGL plots each voxel as an individual point and there are typically 368 x 272 x 256 voxels = 26 Million voxels in the BW image alone, the screen update time for user requests to rotate or zoom the view can take several seconds, making quick inspections of data difficult. To accommodate faster viewing, the application was designed to operate in Cloud mode, where only the voxels representing the outside surfaces of the image volume are plotted to provide a smooth and real time response to
63


the user. The surface clouds of the BW, Red Doppler and Blue Doppler are shown in Fig. Ull.
Figure 11.11 White BW Scan Envelope Overlaid with Red and Blue Doppler
A further benefit to cloud mode is the ability see through or past neighboring voxels for a more intuitive view of the 3D data geometry, particularly for structures that intertwine or obscure each other.
To enable the user to view all of the data points in the selected image volume a raw data viewing mode can be invoked with the RAW button. To provide some level of transparency in raw data mode, the points plotted for each voxel are slightly smaller than the voxel volume to allow the user to see though and past some number of surface voxels to gain insights into the texture of the interior data as can be seen in Fig. II. 12 for the BW
64


volume (where the compound curvature of the 3D transducer can be seen at the top of the image), and in Fig. IF 13 for the Red and Blue Doppler volumes.
Figure 11.12 Raw BW Data Volume (Semi-transparent)
Figure 11.13 Red and Blue Doppler Data in Raw Mode
65


In Cloud mode, the voxels are plotted as saturated (color on/off) false RGB colors for visibility. In Raw mode, the voxels are plotted in linearly mapped (min voxel data value = no color, max voxel data value = max color) false RGB colors relating low voxel value to low saturation hue and high voxel value to high saturation hue as a means of providing the user with relative velocity information and better visibility to low intensity voxels. RawTC mode presents the image data as it would be seen on the iU22 LCD using the Philips velocity/color map. LCD displays implement an additive color scheme where all colors off displays the only the white from the florescent or electroluminescent backlight (white light passed), and all colors on deepens their hues to present a black display (white light blocked). The RawTC color mapping transitions from black surrounding the vessel to a dark hue of the color representing low velocity at the vessel wall and then through lighter hues of the color to near white for the highest velocity at the center of vessels. The hue/velocity variation in the vessel cross-section shown in Fig 11.14 suggests a Poiseuille (parabolic) velocity profile, where Plug (uniform) velocity profile would be indicated by a uniform color hue across the vessel cross-section.
66


Figure 11.14 Red and Blue Doppler shown in RawTC (True Color) Mode
Importing Data
Ultrasound images captured on the iU22 are transferred to the Laptop PC running the VAS development environment using a USB memory stick. The image file is opened by a conversion program developed in MATLAB which opens the file using the Philips (*.p) library function and then exports the data as four separate files: three binary image arrays (BW, RED, BLUE), and a text (*.txt) file which contains the x, y, z image dimensions and floating point voxel dimensions. These files are imported by VAS for use by the developer and clinician.
67


Capturing a Region of Interest (ROI)
After the clinician selects an image to import, VAS scans each data file and creates the surface point cloud of the Doppler data after which it is available for display and ROI selection in the main window.
The ROI Set, ROI Clear, and ROI View functions allow the clinician to manually select the ROI. The clinician can set the bounding box using a Size and Position method, or, by setting the Upper Left, Lower Right, Left and Right sides of the ROI box, or, as a future enhancement, by Dragging the box size and position. The 3D corners of the box are determined by the intersection of boxes projected from two orthogonal views. The clinician first sets the ROI box size and position in the lateral view, the view is then switched to Elevation view where the box is sized and positioned to complete the 3D corner selection. A click of the right mouse button steps the clinician through each selection, or, the GUI buttons can be used to direct the selections and completion.
VAS was implemented in 7500 single spaced lines of original code.
Aim 3 Autonomously Segment Vasculature and Compute Centerline
Segmentation
The initial method developed for finding the external surfaces of the Doppler data was a raster scan and threshold edge detection algorithm that was implemented within the function responsible for the initial reading of the Doppler data file for processing efficiency. This method was used in both ADP and VAS.
68


ADP Segmentation
To capture circular and slightly elliptical cross-sections of sliced vessels and not vessels cut along their length, a Masking (spatial filtering) function was developed and invoked by the MSK button on the ADP GUI. As ultrasound image data contains spatial noise (volumetric blobs unrelated to vessels) which would interfere with the vascular surface reconstruction, the MSK process was assisted by the Noise Reduction (NR) function which was developed to eliminate cross-sectional objects with an area below a size set by the slider control labeled NR using the MATLAB bwareaopen and imfill library functions. After an image is processed by MSK and NR, a Contour (CNT) function identifies the perimeter of the mask identified objects which are plotted as a blue overlay on the cross-section; a Center (CEN) function computes the objects center overlaid as a red circle; and an elliptical representation of the object is overlaid as a red ellipse. These functions utilize the MATLAB contour, bwconncomp and regionprops library functions.
Smoothing of the Color Doppler surface was implemented in a function invoked by the (SUR) which utilizes the MATLAB smooth3 library function and a 9x9x9 voxel Gaussian convolution kernel. The smooth3 function convolves each voxel of the Color Doppler data volume with the kernel that has a Gaussian weighting of one at the center voxel and tapers the values of the surrounding voxels in the cube to smaller numerical values at the edge of the kernel cube following a Gaussian distribution. This process effectively shifts voxels from their original positions to nearby positions to create a smoother spatial surface.
69


VAS (C/OpenGL) Segmentation [Visual Human Data]
TolTech (Aurora, CO) provided an STL/OBJ file format mesh for the portal vein structure that was used to fabricate the flow phantoms described in Chapter IV [58], The mesh provided was created using matrix methods that applied uniform threshold rules to the entire image data set, and as such, was not able to re-create the more subtle details of local vessel geometry. The provided mesh also contained many fragments and holes which required considerable manual smoothing as described in Chapter IV [58],
However, the database for the provided mesh was also the basis of TolTechs commercial anatomy training software VHDissector Pro, which provides high resolution [1760 x 1024] head-to-foot photographic images of slices of a human male with manually segmented hepatic vasculature used for medical training. For improved hepatic modeling, the desired hepatic vascular structures were extracted from these images by activating the segmentation detail overlay for the portal vein, hepatic vein, and liver capsule individually, and then manually saving each of the 180 superior to inferior axial slices comprising the liver volume as individual BMP image files.
As the manual vessel segmentations appeared as purple/turquoise color overlays in the images, Adobe Photoshop was used to identify the color gamut values of the component Red/Green/Blue (RGB) hues such that the segmentation could be discriminated from the general yellow/brown anatomy.
The C/OpenGL version of VAS was written to open each image file, capture the dimensions and shape of each vessel cross-section, and then display in 3D virtual reality the resulting point cloud for the user to view. The serial data stream for each
70


image was read from the BMP file as a horizontal then vertical image raster scan of voxels (pixels in the 2D plane). The incoming voxel values were compared to the RGB color range of the Visual Human segmentation color. For voxel values within the RGB color criteria, the leading and trailing voxels positions of the vessel cross-section were recorded and then the halfway point between them was computed and recorded. This process created a list of line segments where the endpoints represented the voxels of the surface point cloud.
Vessel Edge Detect Pseudo-code
for all Visual Human BMP files (n=180) for all horizontal raster lines in the BMP file for all voxels in the current raster line for each line segment meeting segmentation color rules record the first voxel position record the last voxel position
compute the difference between the first and last voxel as the segment center
Line segment centers were used to differentiate vessel cross-sections in close proximity since line segment centers are more central to the vessel cross-section and thus further away from each other than the edge voxels of two adjacent vessel cross-sections.
Vessel Object Pseudo-code
for each slice in the image volume (n=180) for all the line segments in the slice
find the line segments centers that are within q voxels of each other and assign them to a unique vessel cross-section object
71


The center of vessel cross-sections were calculated using an average center of
mass method.
Vessel Center Pseudo-code
for all slices in the image volume (n=180) for all vessel cross-section objects for all the line segments in the object sum the x values of the segment centers sum the y values of the segment centers
Divide the final x andy sums by the number of segment centers to get the average x andy location of the mass center
The gross dimensions of the vessel cross-sections were detected and saved. The segmentation function identifies the edges of the vessel (ends of line segments) with red dots, line segment centers with green dots, the mass centroid as a single white dot, and the maximum vessel dimensions with a grey bounding box.
Further C functions were written using the OpenGL graphics language to create a viewing window and keyboard based controls so that the clinician can view and manipulate the point cloud of the vascular structure as a 3D object in a virtual world. The portal vein structure is displayed in red, the hepatic vein in green, and the liver capsule in blue. As the clinician changes the point of view, the integer voxel positions in the static array holding the image data are translated into floating point values using the physical voxel sizes provided by the ultrasound cart and are then plotted in the OpenGL viewing volume.
72


Total Volumes
A function was created to sum the volume of the voxels captured with within the portal vein, hepatic veins and liver capsule using the floating point voxel dimensions.
Volume Pseudo-code
for all the slices in the image volume (n=180) for all the line segments in the slice
compute the number of voxels between the line segment end points sum the voxel count
multiply the sum by the voxel dimensions to get total volume
CAD File Export
A function was created to export the voxels representing the portal vein, hepatic vein, and liver capsule surface point clouds in a standard (*.xyz) file format for importing to other Computer Aided Design (CAD) software.
VAS (C/C++ /OpenGL/Qt) [Ultrasounddata]
In the C/C++/OpenGL/Qt version of VAS, a nearest neighbor/infection
method of identifying and tagging vessel segments was developed. Each vessel segment
identified was color coded for ease of viewing. Each voxel captured in the ROI is tested
for association to an object, where voxels near voxels already associated with an object
are assigned to the same object, and voxels not near assigned (infected) voxels are
assigned a new object number. The ROI volume is scanned systematically to ensure all
voxels are assigned and color coded by object number for identification when viewed.
The spatial capture range (infection distance) can be adjusted to accommodate varied
types of noise and close proximity of the vessels. Use of small capture ranges is good for
73


discriminating close objects at the risk of an object being fragmented into multiple objects if it contains internal voids which are larger than the capture range.
Vessel Segmentation Pseudo-code
for all voxels in the ROI
test for association to an object within an adjustable capture range if voxel not already assigned if there is a voxel assigned within the capture range assign the object number else
assign a new object number
Centerline
ADP Centerline
As part of the Contour (CNT) function, the MATLAB regionprops library function was called to identify the center of vessel cross-sections found by the segmentation method.
VAS Centerline
A special centerline finding algorithm was developed using a mass centroid approach to create line segments which approximate the centerline of each vessel captured within the ROI. The algorithm partitions the ROI into an adjustable number of compartments defined by planes perpendicular to the beam direction. A mass centroid algorithm is used to identify the center of mass of the compartment.
74


Centerline Pseudo-code
divide the ROI into n compartments for all compartments for all voxels within the compartment
sum the x values of the voxels in the compartment sum the y values of the voxels in the compartment sum the z values of the voxels in the compartment Divide the final sums by the number of voxels in the compartment to get the x,y, z locations of the mass center
Line segments are drawn graphically between the compartment centers to create the user visible centerline.
Aim 4 Autonomously Compute Segmented Vessel Diameters and Cross-sectional Areas
Diameter
ADP Diameter
A function was written to calculate the vessel cross-sectional diameter using the MATLAB image processing function regionprops.
VAS Diameter
Equation II.3 shows that any voxel B forming a vector CB with voxel C on the centerline which satisfies the Dot Product rule of AC CB = 0 with the vessel centerline AC within a set tolerance, lies on the cross-section of the vessel and defines a vessel radius of CB which is half the vessel diameter.
75


A-
-
ACCB = 0
Equation II.3 Voxels in the Vessel Cross-sectional Plane
Once the on-plane cross-sectional voxels are located, the distance from the centerline to each voxel is located and the furthest three voxels were averaged to provide a radius for use in the vessel cross sectional area calculation described below. The radius averaging accommodates a small amount of eccentricity in the cross-section.
Diameter Pseudo-code
for two centerline segment points on the centerline (top and bottom ofROI) for all voxels within the ROI
if voxel satisfies dot product within tolerance of +/-0.0001 tag voxel as being on cross-section
compute distance from centerline to voxel and store the 3 longest average the longest three radii and double to get diameter
The estimated diameters were compared to the phantom tubing inner diameter to calculate the percentage error as (Dest Dref)/Dref so that larger than reference diameters would give a positive error.
76


Area
ADP Area
A function was written to calculate the vessel cross-sectional area using the MATLAB regionprops library function.
VAS Area
For speed and ease of calculation, the geometric form for the area of a circle based on radius (Eq. II.4) was implemented.
Vessel Area = nr2
Equation II.4 Area of Vessel Cross-section
Aim 5 Autonomously Compute Angle Corrected Velocities for Segmented Vessels
Velocity
ADP Velocity
Velocity algorithms were not developed on the ADP as all development migrated to VAS.
VAS Velocity
The velocity of the fluid captured within each voxel is provided as a value in the decimal range of 0-255. Since there may be many voxels (points of unique velocity) in the cross-sectional capture, the velocities of all of the voxels captured are averaged to
77


create a mean for the cross-section as shown in Eq. II. 5 where n is the number of voxels in the cross-section.
Yd voxel value (Color Doppler velocity)
Mean Velocity =---------------------------------------------
Equation II.5 Cross-sectional Mean Velocity
Use of cross-sectional mean velocity is important when Poiseuille (parabolic) flow is present since the velocity at the center of the flow may be much greater than the flow at the walls of the vessel where the velocity is zero. As discussed in Chapter I, Poiseuille flow is found downstream of branching points and bends where the fully developed flow takes on a parabolic velocity profile. The use of a cross-sectional mean velocity calculation for a plug flow profile is not an issue as it is already a constant. Centerline velocity values were obtained from individual voxels along the centerline.
VAS Angle Corrected Velocity
As discussed previously, computing accurate flow in the direction of the vessel centerline requires an angle correction of the Doppler velocity data. To automate the process of angle correcting vessels captured in the ROI along the centerlines of captured vessels, an algorithm was developed to automatically determine the focal points of the transducer fan beam by detecting the edges of the BW data cloud, computing a line fit along the edge lengths, and displaying the points of intersection. Spatial filtering was implemented to detect the noisy BW fan beam envelope edges accurately. The BW
78


volume edges are located by sampling the BW data volume for non-zero voxels at five levels in the depth axis by propagating a plane of 40 x 40 voxels halfway into the BW data volume in the lateral direction. If no voxels are touched by the plane, the plane is stepped in the elevation direction and the process repeated until voxels are contacted. As soon as an edge is detected the algorithm moves on to the next sample point in depth. A line is then drawn through the points of edge contact and projected towards the transducer and is mirrored to the other three comers in two steps to provide intersections representing the lateral (phased array) sweep origin and the elevation (mechanical) sweep origin. The 2D projection of the lateral sweep origin is used as the transducer origin as that is the origin used for the 2D Power Doppler angle correction on the ultrasound cart.
For autonomous angle corrected velocity, the two end segments (ROI inlet and outlet) of the computed centerline are used. The Cosine of the angle between the insonation beam from the transducer origin to one end of each centerline segment is computed using a trigonometric form of the Dot Product relationships (Eq. II. 6).
cos 0 =
AB
11
Vac =
Vdop COS 0
A = vessel direction B = insonation direction
0 = angle between insonation and vessel direction Vdop = voxel-wise Color Doppler velocity Vac = voxel-wise angle corrected velocity
Equation II.6 Angle Between Insonation Beam and Centerline Segment
79


Full Text

PAGE 1

AUTONOMOUS METHODS FOR VESSEL RECONSTRUCTION AND QUANTIFICATION OF VASCULAR FLOW by BRYAN E. YUNKER Certificate, University of Colorado, 2009 Bachelor of Science, Wash ington State University, 1981 A thesis submitted to the Faculty of the Graduate School of the University of Colorado in partial fulfillment of the requirements for the degree of Doctor of Philosophy Bioengineering Program 2014

PAGE 2

ii 2014 BRYAN E. YUNKER ALL RIGHTS RESERVED

PAGE 3

iii This thesis for the Doctor of Philosophy degree by Bryan E. Yunker has been approved for the Bioengineering Program by Robin Shandas, Chair Kendall S. Hunter, Advisor Gerald D. Dodd Yusheng Feng S. James Chen Samuel Chang July 9, 2014

PAGE 4

iv Yunker, Bryan E. (Ph.D., Bioengineering) Autonomous Methods for Vessel Reconstructio n and Quantification of Vascular Flow Thesis directed by Assistant Professor Kendall S. Hunter. ABSTRACT Introduction: The success rate for the Radiofrequenc y (RF) ablation of hepatic tumors is limited by local blood flow carrying away the heat applied during treatment, which can prevent some tumor cells from reaching the ne crosis (cell death) temperature of 55 C. As previous studies [1] suggest that local flow may be used as a predictor of ablation size and device performance, the focus of this study explores a method of determining local flow by reconstructing hepatic vasculature structure from 3D ultrasound. With the hypothesis that the outer surface of 3D Doppler data can accurately represent the inner surface (lumen) of a vessel, two software a pplications (ADP and VAS) were created as platforms for developing and testing algorith ms capable of meeting a clinical Time-toResults (TTR) goal of 10 seconds for local flow estimation. Methods: A portal vein structure was acquired for use as a pattern for flow phantom construction and CFD flow simulations. Industrial casting materials we re imaged with ultrasound, MRI and CT, and then used to fabricate flow phantoms. The flows through the portal phantom branches were measured over a range of flow rates. A Computational Fluid Dynamics simulation was created and compared to the measured phantom branch flows. Algorithms capable of estimating local flow were developed and evaluated using simple tubing flow phantoms. Results/Discussion: Each of the materials test ed for phantoms showed useful

PAGE 5

v characteristics in medical imaging applica tions. Two portal vein flow phantoms were successfully fabricated, one as a solid silicone block, and the other as a thin walled latex tube structure. The portal phantom bran ch flow measurements and the CFD model estimates compared within 17%. Color Doppler surfaces within the phantom lumen could not be obtained with the water based test fluids used, however: testing with Doppler surfaces ~2x larger than reference provided random diameter estimate deviations under 9%, angle corrected veloc ities within 15% of referenc e, and net flows within 15% of reference. The segmentation, centerline, diameter, area, angl e correction, and flow algorithms functioned correc tly and within the TTR. Conclusions: The rapid prototyping fabrication techniques and ma terials employed resulted in usable anatomical flow phantoms. CFD modeling successf ully predicted the performan ce of the anatomical flow phantoms. The algorithms developed for local flow assessment are ready for accuracy testing with a more echogenic test fluid (blood ). The diameter/area algorithm needs to be improved to handle non-circular vascular st ructure. The error performance for all algorithms needs to be improved. The form and content of this abstract are approved. I recommend its publication. Approved: Kendall S. Hunter

PAGE 6

vi DEDICATION I dedicate this work to my parents, gr andparents, family, friends, committee, and colleagues for their unwaver ing support and patience.

PAGE 7

vii ACKNOWLEDGMENTS I would like to thank the following individuals for their support, guidance, and teachings: Committee Members: Kendall S. Hunter Gerald D. Dodd, Yusheng Feng, Robin Shandas, S. James Chen, and Samuel Chang Fellow Authors: Dietmar Cordes, Ann L. Scherzinger Liver Lab Manager: Anthony Lanctot Bioengineering Instructor and Lab Manager: Craig Lanning Lab Assistants: Bryan Rech, Je nnifer Wagner and Luciano Mazzaro Professors: Gary Fullerton, Emily Gibson, Mark Golkowski, William Hendee, Willem Schreder, Mark Brown Fellow Students: Arati Gurung, Shawna Burg ett, Steve Humphries, Mike Zimkowski Funding Sources: NIH/NCI 1F31CA171620-01A1, Colorado Translational Research Imaging Center (C-TRIC); NSF 0932339; University of Colorado, SOM, Department of Radiology; NIH T32HL072738; K25 HL094749; K24081506; RO1HL114753; and NHLBI K25-094749 Equipment: Ged Harrison, Philips Healthcare, Bothell, WA

PAGE 8

viii TABLE OF CONTENTS CHAPTER I. INTRODUCTION ....................................................................................................... 1 Background and Significance ................................................................................. 1 Disease Impact ................................................................................................... 1 Disease Treatment .............................................................................................. 2 Research Objective ............................................................................................. 8 Estimating Local Blood Velocity/Flow .............................................................. 9 Background Studies .............................................................................................. 12 Hepatic Physiology ........................................................................................... 12 Hepatic Flow Rates .......................................................................................... 22 Vascular Segmentation ..................................................................................... 32 Clinical Requirements ...................................................................................... 41 II. RESEARCH .............................................................................................................. 47 Specific Aims ........................................................................................................ 47 Aim 1 Fabricate and Model a Portal Vein Flow Phantom ............................. 47 Aim 2 Obtain an ROI from a 3D Image Volume from Clinician Input ......... 47 Aim 3 Autonomously Segment Vascul ature and Compute Centerline .......... 47 Aim 4 Autonomously Compute Segmented Vessel Diameters and CrossSectional Areas ................................................................................... 48

PAGE 9

ix Aim 5 Autonomously Compute Angle Corrected Velocities for Segmented Vessels ................................................................................................ 48 Aim 6 Autonomously Compute Volumetric Flow for Segmented Vessels ... 48 Materials and Methods .......................................................................................... 49 Aim 1 Fabricate and Model a Portal Vein Flow Phantom ............................. 49 Aim 2 Obtain an ROI from a 3D Image Volume from Clinician Input ......... 57 Aim 3 Autonomously Segment Vascul ature and Compute Centerline .......... 68 Aim 4 Autonomously Compute Segmented Vessel Diameters and CrossSectional Areas ................................................................................... 75 Aim 5 Autonomously Compute Angle Corrected Velocities for Segmented Vessels ................................................................................................ 77 Aim 6 Autonomously Compute Volumetric Flow for Segmented Vessels ... 80 Results and Discussion ......................................................................................... 81 Aim 1 Fabricate and Model a Portal Vein Flow Phantom ............................. 81 Aim 2 Obtain an ROI from a 3D Image Volume from Clinician Input ......... 91 Aim 3 Autonomously Segment Vascul ature and Compute Centerline .......... 93 Aim 4 Autonomously Computed Se gmented Vessel Diameters and CrossSectional Areas ................................................................................. 104 Aim 5 Autonomously Compute Angle Corrected Velocities for Segmented Vessels .............................................................................................. 117 Aim 6 Autonomously Compute Volumetric Flow for Segmented Vessels 137 III. FULL TEXT OF AIM 1 PHANTOM MATERIALS (PUBLISHED) ................. 142 IV. FULL TEXT OF AIM 1 PHANTOM FABRICATION (PUBLISHED) ............. 169 V. CONCLUSIONS ..................................................................................................... 187

PAGE 10

x VI. FUTURE RESEARCH ............................................................................................ 191 REFERENCES ........................................................................................................ 197

PAGE 11

xi LIST OF TABLES TABLE I.1 Normal Portal Vein Size, Fl ow, and Pressure Characteristics .................................. 22 I.2 Blood Properties ........................................................................................................ 26 I.3 Thermal Characteristics of Blood and Liver ............................................................. 27 I.4 Hepatic Ultrasound Characteristics ........................................................................... 31 I.5 Summary of Vascular Segmentation Approaches ..................................................... 37 II.1 Centerline Error from Surface Noise ..................................................................... 103 II.2 Diameter Error Statistics ........................................................................................ 106 II.3 Diameter Errors from Centerline ............................................................................ 107 II.4 Area Errors from Diameter Error ........................................................................... 117 II.5 Uncorrected Velocity Statistics .............................................................................. 124 II.6 Reynolds Numbers for Tested Images ................................................................... 125 II.7 Straight Tube Velocity Errors (S0324-65B) .......................................................... 129 II.8 Curved Tube Velocity Errors (S0324-65R) ........................................................... 133

PAGE 12

xii II.9 Rough Surface Velocity Errors (S0313-70B) ........................................................ 137 II.10 Net Flow Velocity Errors ..................................................................................... 141

PAGE 13

xiii LIST OF FIGURES FIGURE I.1 Layout of RF Ablation Procedure ............................................................................... 3 I.2 RF Probe Tip Temperature Gradient ........................................................................... 4 I.3 RF Ablation Progression ............................................................................................. 4 I.4 Thermal Injury .......................................................................................................... ... 5 I.5 Ablation Volume Vs Flow ........................................................................................... 7 I.6 Ablation Volume Vs VFI ............................................................................................ 8 I.7 Typical Doppler Color Ultr asound Image of Hepatic Veins ..................................... 10 I.8 Location of the Liver in the Circulatory System ....................................................... 12 I.9 GI Vasculature Feed ing the Portal Vein .................................................................... 13 I.10 Vascular Anatomy of the Liver ............................................................................... 14 I.11 2D Image of Portal Vein and Hepatic Artery .......................................................... 15 I.12 Sections of the Liver ................................................................................................ 1 6 I.13 Hepatic Lobule Structure ......................................................................................... 17

PAGE 14

xiv I.14 Hepatic Lobules Flows ............................................................................................ 18 I.15 Lobule Cross-section ............................................................................................... 19 I.16 Detail of the Blood Flow through the Hepatic Sinusoids ........................................ 19 I.17 Hepatocyte Cellular Structure ................................................................................. 21 I.18 Operating Inlet Pressure of the Liver ...................................................................... 23 I.19 Diagram of the Hepatic Flows and Pressures .......................................................... 24 I.20 Viscoelastic Properties of Blood at 2 Hz and 22 C ................................................ 26 I.21 Cirrhosis............................................................................................................... .... 30 I.22 Geometry of Doppler ............................................................................................... 31 I.23 Example of Manual Angle Correction .................................................................... 32 I.24 Hepatic Branching .................................................................................................. 33 I.25 Arterial Phase CT .................................................................................................... 3 9 I.26 Arterial Phase Hepatic Artery .................................................................................. 39 I.27 Venous Phase CT..................................................................................................... 40 I.28 Portal Vein ............................................................................................................ ... 40 II.1 Solid Silicone Body Portal Phantom and Test Setup ............................................... 50

PAGE 15

xv II.2 Thin Walled Latex Portal Phantom and Test Setup ................................................. 51 II.3 Thin Walled Latex Portal Phantom Test Setup ........................................................ 52 II.4 Thin Tube Phantoms ................................................................................................ 54 II.5 BW and Doppler Slices (BW Mode) ........................................................................ 58 II.6 Geometry of Doppler (False Color Mode) ............................................................... 59 II.7 Red and Blue Doppler (False Color) ........................................................................ 60 II.8 Histograms of Red and Blue Doppler ...................................................................... 61 II.9 Linear Color Mappi ng of BW and Doppler ............................................................. 61 II.10 “VAS” Clinical Applicati on Graphical User Interface .......................................... 62 II.11 White BW Scan Envelope Over laid with Red and Blue Doppler .......................... 64 II.12 Raw BW Data Volume (Semi-transparent) ............................................................ 65 II.13 Red and Blue Doppler Data in Raw Mode ............................................................. 65 II.14 Red and Blue Doppler show n in RawTC (True Color) Mode ............................... 67 II.15 Phantom Test Samples ........................................................................................... 81 II.16 Solid Silicone Body Portal Vein Phantom ............................................................. 82 II.17 Thin Walled Dipped Latex Portal Vein Phantom .................................................. 83

PAGE 16

xvi II.18 Solid Body Silicone Portal Vein Phantom ............................................................. 84 II.19 Thin Walled Latex Portal Vein Phantom ............................................................... 85 II.20 Portal Inlet Flow Vs Pressure ................................................................................. 86 II.21 Portal Branch Flow Rates ....................................................................................... 87 II.22 Major Portal Branch Flow Rates ............................................................................ 88 II.23 Velocity Profile Simulati on of Phantom Pattern Mesh .......................................... 89 II.24 Comparison of CFD Estimates with Phantom Measurements ............................... 90 II.25 ROI Selection Box (Lateral View) ......................................................................... 91 II.26 ROI Selection Box (Elevation View) ..................................................................... 92 II.27 ROI Selection Completed....................................................................................... 93 II.28 Edge, Center, and Elliptical Appr oximations of Vessel Cross-section .................. 94 II.29 Gaussian Smoothing of Doppler Data .................................................................... 94 II.30 Outline and Center Comput ed for Vessel Cross-section ........................................ 95 II.31 Typical Vessel Overlap and Proximity .................................................................. 96 II.32 Segmented Point Cloud of Hepatic Capsule and Portal/Hepatic Veins ................. 96 II.33 Centerline Approximation ...................................................................................... 97

PAGE 17

xvii II.34 Multi-Vessel ROI Capture ..................................................................................... 98 II.35 Surface Noise in Color Doppler Point Clouds ....................................................... 99 II.36 Long Curved Vessel Test (S0324-65R) ............................................................... 101 II.37 Centerline Direction Error Due to Surface Noise ................................................ 102 II.38 Color Doppler Cross-se ction Capture (S0324-65B-3) ......................................... 104 II.39 Color Doppler Cross-se ction Capture (S0313-70B-5) ......................................... 104 II.40 Color Doppler Diameter Measurements .............................................................. 105 II.41 Color Doppler Diameter Error Eccentricity (S0313-70) ................................... 108 II.42 Color Doppler Diameter Error Pattern (S0320-34) ........................................... 110 II.43 Surface Detection Threshold Set to 1/128 and 25/128 ......................................... 112 II.44 End Tapering of Color Doppler Surfaces (S0324-65BR) .................................... 114 II.45 Power Doppler Diameter Measurement Tubing (S0407-41) ............................. 114 II.46 BW Image Envelope and Computed Transducer Origins .................................... 118 II.47 Transducer Beam Lines ........................................................................................ 119 II.48 Comparison of Mass/Time and Manual Angle Correction .................................. 120 II.49 Uncorrected Velocities ......................................................................................... 122

PAGE 18

xviii II.50 Location of Velocity Samples .............................................................................. 123 II.51 Manual Angle Correcti on and Color Scale (Cart) ................................................ 126 II.52 Manual Angle Correcti on and Color Scale (VAS) ............................................... 126 II.53 Doppler Velocity True Color Variation (S0324-65) ............................................ 127 II.54 Straight Tube Velocity Errors (S0324-65B) ........................................................ 128 II.55 Straight Tube Blue Top View (S0324-65B) ...................................................... 130 II.56 Curved Tube Top View (S0324-65R)................................................................ 132 II.57 Curved Tube Velocity Errors (S0324-65R) ......................................................... 133 II.58 Series Phantom All Tubes .................................................................................. 135 II.59 Rough Surface Capture Top View (S0313-70B-5) ............................................ 136 II.60 Rough Surface Velocity Errors (S0313-70B) ...................................................... 136 II.61 Net Flow Phantom, Straight-Curved Tubing (S0324-65) .................................... 138 II.62 Net Flow Errors .................................................................................................... 14 0 VI.1 Extracted Vasculature Regist ered to Patient Physiology ...................................... 194 VI.2 Virtual Reality Demo ............................................................................................ 195

PAGE 19

xix LIST OF EQUATIONS EQUATION I.1 Velocity Flow Index (VFI) .......................................................................................... 7 I.2 Flow from Area and Velocity ...................................................................................... 9 I.3 Hepatic Resistance by Windkessel ............................................................................ 28 I.4 Resistive Index for Hepatic Artery ............................................................................ 28 I.5 Portal Vein Reynolds Number ................................................................................... 29 II.1 Mean Mass to Volume Flow Conversion ................................................................. 53 II.2 Velocity to Volume Flow Conversion ..................................................................... 56 II.3 Voxels in the Vessel Cross-sectional Plane ............................................................. 76 II.4 Area of Vessel Cross-section ................................................................................... 77 II.5 Cross-sectional Mean Velocity ................................................................................ 78 II.6 Angle Between Insonation Beam and Centerline Segment ..................................... 79 II.7 Centerline Error from Noise .................................................................................... 103 II.8 Centerline Error ....................................................................................................... 107 II.9 Elliptical Area ......................................................................................................... 116

PAGE 20

xx II.10 Test for Laminar Vs Turbulent Flow .................................................................... 125

PAGE 21

xxi LIST OF ABBREVIATIONS ADP Algorithm Development Platform BW Black and White CT Computed Tomography GUI Graphical User Interface HCC Hepatocellular Carcinoma MRI Magnetic Resonance Imaging RF Radiofrequency RI Resistive Index ROI Region of Interest TTR Time-To-Results VAS Vascular Analysis System VFI Vascularization Flow Index

PAGE 22

1 CHAPTER I INTRODUCTION Background and Significance Disease Impact The American Cancer Society (ACS) he patocellular carcinoma (HCC) and bile duct cancer estimates for 2014 are 33,190 new diagnoses and 23,000 deaths in the United States, where men will lead women 3:1 in diagno sis and 2:1 in deaths [2]. According to the NCI SEER report for 2003-2009, the survival rate from hepatic tumors that have spread to other parts of the body is 2%, and, th e survival rate from local tumors is 28% [3]. Worldwide, the ACS states an av erage of 700,000 diagnoses and 600,000 deaths [2], and Cancer Research UK states 745,000 deaths in 2012 [4]. In addition to HCC diagnoses and deaths over 50% of the individuals diagnosed with colorectal cancer will develop liver metastases, comprising 150,000 diagnoses and 50,000 deaths as of 2008 in the United States [5 ]. Colorectal cancer is the third most common type of cancer at 1.4 million cases annually worldwide, suggesting 75,000 more cases of liver cancer in the United Stat es and 700,000 more cases worldwide beyond primary HCC statistics. In 2013, liver cancer was the fifth most co mmon cause of cancer death in males in the US across all ages [6].

PAGE 23

2 Disease Treatment Systemic chemotherapy is an option for tumor treatment but cures are rare. Surgery offers better survival rates and a longer disease-free period, however, this applies only to a very limited patient population, and, there can be significant postoperative morbidity and high tumor recurrence. Ablati on Therapies can be used to destroy the viability of non-resectable tumor cells incl uding Cryogenic Ablation with liquid nitrogen, Chemical Ablation with a toxin such as Ethanol, and Thermal Ablation with Radio Frequency or Microwave energy. Therapeutic Focus RF Ablation (RFA) was chosen for this st udy as it is a less invasive therapeutic technique that can be used repeatedly to treat inevitable tumor recurrences and it is a therapy performed by the clinical sponsor of this research. There are a number of benefits of RFA [7], including ‘local cure rates [complete ablation] can reach 83% for tumors (<3 cm) with minimal side effects, a major complication rate of only 1-3%, and, patients can go home the same day or after an overnight hospital stay’. Where RF ablation has been shown to be effective for the treatment of small (< 3 cm) malignant hepatic tumors, its utility in treating la rger tumors is quite variable [8].

PAGE 24

3 RFA Technology RF Ablation uses an alternating electric current generator ope rated at 500 kHz, a needle electrode, and large adhesive ground pa ds placed under the patient as depicted in Fig. I.1. Figure I.1 Layout of RF Ablation Procedure [9] When the RF generator is turned on, the concentrated alternat ing electric current around the needle electrode causes ionic agitat ion and frictional heat in the surrounding tissue. An elevation in local temperature hi gher than 55 C causes coagulation necrosis and cell death in tissue and tumors. The therma l injury created by a st raight rod probe is elliptical in shape. The pr obe tip must operate at temper atures over 120 C in order to heat to the entire tumor volume to the necr osis temperature as shown in Fig. I.2.

PAGE 25

4 Covidien (Valley Labs) Figure I.2 RF Probe Tip Temperature Gradient The RF ablation process is shown in Fig. I.3 where the left pane shows the untreated tumor. The middle image shows a thermal injury much larger than the tumor, a margin required to ensure that all of the tumor cells reach necrosis temperature. http://radiology.ucla.edu Figure I.3 RF Ablation Progression

PAGE 26

5 The right image of I.3 shows the post operati ve scar. As the tumor may be larger than the elliptical injury created by a rod pr obe, the tumor must be treated by either many thermal injuries delivered at intervals with in the tumor volume [10] with a single tine probe, or, one thermal injury with a multi-tin ed probe which would cover a larger volume [9] as shown in Fig. I.4. http://radiology.ucla.edu/body.cfm?id=162 Figure I.4 Thermal Injury Technical success of the ablation procedure is determined by whether the entirety of the targeted tumor as well as a circumferential 5-10 mm margin of normal hepatic parenchyma is included in the ablation zone. If a portion of the obvious or microscopic tumor in the immediately adjacent hepatic parenchyma is not included in the ablation zone, local tumor recurrence is inevitable.

PAGE 27

6 RF Device Performance There are several manufactur ers of RF ablation devices and all produce ablations that vary in size between 3-7 cm in diameter. This variation has been unpredictable to date and accounts for the success in treating tu mors less than 3 cm in diameter and the high failure rates seen in tumors greater th an 3 cm in diameter [11, 12]. The dominant factor controlling the volume of coagulation necrosis produced by an RF ablation device is the volume of hepatic blood flow in the region of the tumor [12-14]. Prior work has shown that the size of the ab lation created in a liver by an RF ablation device decreases with increased local blood fl ow [15-24]. In work performed by Chang [1], fresh bovine livers were obtained from a local slaughter house and each liver was perfused by its own blood with heparin added to prev ent coagulation. RF ablations were performed at several sites within the liver microvasculature at se veral blood flow rates and the ablation volume was verified by tissue sectioning and MRI. The blood flows were normalized for liver mass and the data plotted in Fig. I.5 shows a correlation between RF ablation Volume and normalized Blood Flow.

PAGE 28

7 Figure I.5 Ablation Volume Vs Flow [1] [25] An additional relationship was investigated using Color Doppler imagery of the ablation sites and a Volume Flow Index (V FI) computed with Philips QLab software based on Eq. I.1. C = Number of colored pixels in ROI E = Total number of pixels in ROI Equation I.1 Velocity Flow Index (VFI)

PAGE 29

8 A modest correlation was found between ablation size and the computed VFI as shown in Fig. I.6. however, the errors are too large to be clinically usef ul without additional information. Figure I.6 Ablation Volume Vs VFI [1] Research Objective To increase the success rates stated by Lu [7] and to expand the number of eligible patients with larger tumors, the si ze of the ablation zone that will be produced by a specific ablation device in a given patient must be known in advance of the treatment. With this knowledge, the clinician can deve lop an appropriate treatment plan that

PAGE 30

9 includes the position and number of ablations that will need to be performed to produce an ablation zone of sufficient size to erad icate a tumor without excessive margins. As the relationships shown in Figs. I.5-6 suggest that ablation size prediction and improved surgical success are dependent on quantification of local blood flow, the research presented in this dissertat ion attempts to develop a method of quantifying local blood flow. Estimating Local Blood Velocity/Flow Fluid flow in vessels (pipes/tubes) can be obtained from the cross-sectional area of the vessel and the velocity of fluid fl owing through that cross-section using the relationships in Eq. I.2. Equation I.2 Flow from Area and Velocity Vessel cross-sectional area and centerline direction can be obtained from the physical structure of the vessel (shape of the pipe) which can be obtained from the patient using 3D ultrasound. Black and White (BW) ultrasound data can provide structural information from tissue density differences. Color Doppler ultrasound can provide vessel

PAGE 31

10 structure if the image is acquired with no “color blooming” outside the vessel lumen since the outer surface of the color volume woul d represent the inner surface of the vessel as shown in Fig. I.7. Figure I.7 Typical Doppler Color Ul trasound Image of Hepatic Veins The velocity of the blood in the vessels can be obtained from the Color Doppler velocities if they are Angle Co rrected for the vessel direction relative to the transducer beam direction.

PAGE 32

11 Challenges/Risks The customary procedure for ensuring that Color Doppler and Power Doppler data volumes represent the interior volume of a vessel involves: 1) adjusting the ultrasound power and 2D BW gain to levels that show vessel lumens that are visible enough to be seen but do not vary visibly in position or size with excess signal power or gain; 2) increasing the Color or Power Doppler until the color just fills the lumen interior. The accuracy of these adjustments are inhere ntly variable between ultrasound carts and operator judgment and skill. Success also depends on sufficient echogenicity from the blood to reflect enough sound energy for the rendering of a luminary surface. Hypothesis: Localized patient specific blood flow can be accurately estimated from hepatic 3D Color Doppler data volumes

PAGE 33

12 Background Studies Background studies were performed in th e areas of hepatic physiology, hepatic blood flow, blood properties, hepatic ultrasound and vessel reconstruction in order to develop aims supporting the hypothesis. Hepatic Physiology The location of the liver in the circ ulatory system is shown in Fig. I.8. Figure I.8 Location of the Li ver in the Circulatory System

PAGE 34

13 The liver takes in low pressure non -pulsatile venous blood from the Gastrointestinal tract as shown in Fig. I. 9 and passes it back to the heart after rebalancing the blood chemistry and stori ng energy for use during sleep. Figure I.9 GI Vasculature Feeding the Portal Vein

PAGE 35

14 The anatomy of the liver is shown in Fig. I.10 where the incoming venous blood enters the liver through the porta l vein which branches down to the capillary level to pass the blood through the liver ce lls (Hepatocytes). Figure I.10 Vascular Anatomy of the Liver

PAGE 36

15 The hepatic veins of the liver collect the blood processed by the hepatocytes and return the blood to the hear t through the inferior Vena Cava. The Hepatic Artery provides oxygenated blood to the liver to supplement the de-oxygenated blood coming from the GI tract. The liver also creates bile for storage in the Gall Bladder which is used to emulsify fats in the GI tract. Figure I.11 is an X-ray image of the liver s howing the internal va scular structure. https://research.brown.edu/images/pimages/1232327670.jpg Figure I.11 2D Image of Port al Vein and Hepatic Artery

PAGE 37

16 The surgical regions of the liver are show n in Fig. I.12. as divided into eight volumes that follow the major branches of the vasculature and allow volumes containing malignancies to be resected without damaging the organ. The liver can completely regrow from 25% of its original structure and mass. http://www.cpmc.org/advanced /liver/patients/topics/liv er-cancer-profile.html Figure I.12 Sections of the Liver

PAGE 38

17 The primary cells of the liver (Hepat ocytes) are organized into repeating structures of six sided ‘lobules” as show n in Figs. I.13-14 where the organ level “plumbing” of the portal vein, hepatic arte ry, bile duct, and hepatic veins are shown scaling down to the cellular level. https://courses.stu.qmul.ac.uk/SMD/Kb/m icroanatomy/d/alimentary/index.htm Figure I.13 Hepatic Lobule Structure

PAGE 39

18 http://www.aviva.co.uk/library/images/me d_encyclopedia/cfhg5 76livgal_005.gif Figure I.14 Hepatic Lobules Flows The portal vein, hepatic artery and bile ducts follow the edges of the lobules and are referred to as the “Triad”. The portal vein and hepatic artery blood flow from the triad through “sinusoids” lined with hepatocytes to reach the central veins (acini) that are attached to the hepatic vein as shown in Figs I.15-16. Bile created by the hepatocytes is removed though the “bile canaliculi”.

PAGE 40

19 http://www.ouhsc.edu/histolog y/Glass%20slides/88_01.jpg Figure I.15 Lobule Cross-section www.nature.com Figure I.16 Detail of the Blood Fl ow through the Hepatic Sinusoids

PAGE 41

20 The hepatocytes nearest the portal vein inlet from the triad are most likely to be damaged by toxins (accidental acetaminophen overdose) or de-oxygenation. There are other cells in the liver with specialized functi ons. Stellate cells (Ito) store fat and Vitamin A and produce collagen. Kupffer cells ar e specialized macrophages which remove bacteria and debris from the blood. Dendrit ic cells are immune cells that process antigens. Hepatocytes are 30-40 um wide and line the walls of the sinusoid. A schematic of the cell is shown in Fig. I.17.

PAGE 42

21 Figure I.17 Hepatocyte Cellular Structure In summary, the liver is situated in a major blood collection path to enable the cleansing of a continuous fracti on of venous blood and the st orage of energy. The large size of the liver facil itates a low blood flow and low pre ssure drop which ensures that the hepatocytes have enough time to in teract with the passing blood.

PAGE 43

22 Hepatic Flow Rates To gain a better understanding of normal hepa tic flow rates a literature search was performed and the findings are compiled in Table I.1. Table I.1 Normal Portal Vein Size, Flow, and Pressure Characteristics Flow (ml/min) Pressure (mmHg) Diameter (mm) Area (cm2) Mean velocity (cm/s) # Age M/F Ref min mean max min max min mean max mean min mean max 899 0.99 15.3 85 [26] 724.2 12 4.524 14.2 10 21-55 f [27] 1000 1200 6 [28] 9.8 10.5 11.2 15.7 21.8 27.9 19 20-40 m [29] 7.7 8.3 8.9 26.3 32.5 38.6 18 20-40 f [29] 9.6 43 adult m/f [29] 8.9 36 22-75 m/f [29] 9.4 37 20-40 m/f [29] 11.7 30 24-63 m [29] [30] 11 30 41-75 m/f [29] [31] 847 1.27 11.08 16 18-58 m [32] 715 1 12.43 13 18-58 m [32] 878 1.01 14.4 19 21-58 f [32] 613 0.75 15.26 14 21-58 f [32] 5 8.5 12 21 0-10 m/f [33] 7 10 13 20 11-20 m/f [33] 6 11 15 49 21-30 m/f [33] 6 11 15 58 31-40 m/f [33] 11 107 21-40 m/f [33] 13.6 32.15 55 29 23-65 m/f [34] 7 12 [35] The flows are reported for the portal vein and do not include hepatic arterial flows. The data in the table reflects that di ameter and velocity are easily obtained from non-invasive ultrasound measurem ents using caliper tools in 2D and Pulse Wave Doppler ultrasound. Pressures are much more difficu lt to obtain as they cannot be directly measured in human subjects and must be inferred by techniques such as the Hepatic Venous Pressure Gradient (HVPG) [36], explaining the sparse da ta entries in the table.

PAGE 44

23 The hepatic pressures reported in Table I.1 are in relative agreement with the Venous Pressures suggested by the annot ation in Fig. I.18. at 5-7 mmHg. Figure I.18 Operating Inlet Pressure of the Liver

PAGE 45

24 The normal hepatic operating ranges are shown together in the diagram of Fig. I.19. Figure I.19 Diagram of the Hepa tic Flows and Pressures [28] The hepatic veins collect the sum of th e portal and hepatic artery flows (1200 + 400 = 1600 ml/min) and as the adult liver vol ume is roughly 1600 ml, the liver blood is changed about once per minute. As the a dult human body contains roughly 5 L of blood (4.57L F/5.32L M) [37], the liver could theoretically process the entire body blood volume within 5 minutes, however, it takes mu ch longer since the liver is only one of many venous returns to the heart and the GI tract is only one of many r ecipients of aortic blood. There is a small pressure drop from 7 mmHg at the portal inlet to 4 mmHg at the hepatic vein connection to the inferior vena cava [24]. The portal flow is non-pulsatile having come from the venous side of the GI tract, whereas the hepatic artery flow is highly pulsatile as it branches from the descen ding aorta just below the heart. Heat added

PAGE 46

25 to a liver tumor during the RF ablation rais es the temperature of the surrounding tissue and local blood. Estimating this heat distribution will be treated in follow-on research. Blood Properties Blood consists of plasma (55%), red blood cells (44%), and white blood cells by volume (<1%). Hematocrit is a measure of the red blood cell count to the total blood volume and has a value near 45%. As plasma is primarily water (92%) and contains a low content of albumen (7%) [38] [Red Cross] it is considered a Newtonian (ideal) fluid where the shear stress ( ) and strain rate (d /dt) are proportional, or, /(d /dt) = 1. As whole blood contains deformable red blood cells and other cells it ex hibits viscoelastic properties depending on the vessel size. Viscoela sticity refers to the ability of the 2x7 um dumbbell shaped red blood cells to deform el astically to fit through capillaries, thus changing the blood viscosity as shown in Fig. I.20.

PAGE 47

26 Vilastic Scientific Figure I.20 Viscoelastic Propertie s of Blood at 2 Hz and 22 C In vessel diameters above 0.5 mm, whole blood behaves as a Newtonian fluid and below that diameter whole blood behaves as a non-Newtonian “shear thinning” fluid where the shear stress exceeds the strain rate, /(d /dt) > 1. The dynamic viscosity of whole blood also varies with temperature a nd hematocrit [39] and measured values are provided in Table I.2. Table I.2 Blood Properties Substance Dynamic Viscosity ( ) (kg/ms) Density kg/m3 Conditions Ref Blood 4E-3 1060 37C 45% hematocrit [39] Blood 1058 [40] Water 1E-3 [38] [39] Water 992.2 40C [various]

PAGE 48

27 The thermal characteristics of blood are shown in Table I.3. Table I.3 Thermal Characte ristics of Blood and Liver Substance/Tissue Thermal Conductivity (W/mK) H20 (%) Aortic .476 54 Whole .492 54 Plasma .570 93 Pure Water .627 100 Liver .469/.564 77 Tumor peripheral .511 Tumor core .561 Hepatic Flow Dynamics Hepatic Resistance As illustrated in the radiograph of Fig. I .11, the liver distributes portal vein inflow to the hepatic sinusoids at the capillary/cellular level and the hepatic veins re-collect the flows for return to the inferior vena cava. While modeling the resistance of flow at the branching and sinusoidal levels would be very complex, a rough figure of resistance can be obtained by treating the liver as a c ontrol/storage volume and using a reduced form of the Windkessel equation (Eq. I.3) with the following assumptions; portal flow is 75% of the total in-flow, hepatic flow is 25% of the inflow, inflow equals outflow after the liver is full of blood, steady flow, no cha nge in liver volume, portal pressure of 7 mmHg (933.25 Pa), outlet pressure of 4 mmH g (533.29 Pa), and from Table I.1 an average flow rate of 800 ml/min (13E-6 m3/s).

PAGE 49

28 Equation I.3 Hepatic Resistance by Windkessel In clinical practice, a Resistive Index (RI) is computed from the systolic and diastolic velocities of the hepatic artery using Eq. I.4. Equation I.4 Resistive Index for Hepatic Artery Hepatic Portal Vein Flow Profile Modeling the portal vein as a simple smooth pipe flow, the Reynolds number provided by Eq. I.5 is an indi cator of laminar (Re < 2300 ) vs turbulent (Re > 6000) flow [41]. Laminar flows take on a parabolic velo city profile across the vessel cross-section

PAGE 50

29 downstream from curves and stenosis (Poi seuille), where turbul ent flows exhibit a uniform velocity profile (Plug). Using the density of blood of 1060 kg/m3 and a dynamic viscosity of 4E-3 kg/ms from Table I.2, and, an average portal vein diameter of 10 mm (.01 m) and velocity of 15 cm/s (0.15 m/s) from Table I.1. Equation I.5 Portal Vein Reynolds Number The result from Eq. I.5 is well below the approximate threshold for laminar flow suggesting that the portal vein flow is laminar. Pathological Flow The principle inhibitor to flow through th e liver would be a diagnosis of cirrhosis, a fibrosis scarring around he patocytes which have been injured or killed by disease (Hepatitis B) or toxins (acetaminophen) th at increases the resistance to blood flow through the damaged sinusoids as shown in Fig. I.21.

PAGE 51

30 Figure I.21 Cirrhosis This increased resistance to flow increases the pressure drop across the liver and can lead to a condition of Portal Hypertension which is clinically indicated by a portal vein pressure greater than 12 mmHg [31] or a por tal velocity less than 16 cm/s [28]. The velocity indicator is not conclusive alone as it overlaps the normal ranges listed in the studies of Table I.1. Hepatic Ultrasound As ultrasound is a low cost imaging modality and a 3D system was available for this project, the characteristics of liver a nd surrounding tissues was compiled in Table I.4. Fibrosis

PAGE 52

31 Table I.4 Hepatic Ultrasound Characteristics As Color Doppler velocities are relativ e to the origin and direction of the ultrasound beam, it is necessary to perform an angle correctio n between the direction of the vessel centerline and the direction of insonation as depicted in Fig. I.22. Figure I.22 Geometry of Doppler Tissue Density (kg/m3) c (mm/uS) Attenuation (dB/cm) @ 1MHz Impedance (z) (rayls) Ref Liver 1061 1.555 .4 .7 1.65E6 [40] Liver 1060 1.566 1.66E6 [39] Blood 1058 1.560 .18 1.65E6 [40] Blood 1060 1.566 1.66E6 [39] Fat 924 1.450 .5 1.8 1.34E6 [40] Fat 920 1.446 1.33E6 [39] Muscle 1068 1.600 .2 .6 1.71E6 [40] Muscle 1070 1.5421.626 1.651.74E6 [39]

PAGE 53

32 In current practice, the a ngle correction procedure is performed manually on the screen of the ultrasound cart by the clinician fo r each location of inte rest along the vessel as shown in Fig. I.23 by placement of a sample volume box in the center of the vessel and rotating a cursor in-line w ith the vessel flow. Figure I.23 Example of Manual Angle Correction Vascular Segmentation Hepatic Vascular Structure Many images of hepatic vascular structure were viewed to gain insights into the challenges of vascular reconstruction. The best physiological ex ample of the finely arborized hepatic branching is shown in Fig. I.24. User Located Sample Volume and Cursor Set In-line with Flow Do pp ler Beam Directio n Flow Direction

PAGE 54

33 Figure I.24 Hepatic Branching (Google Images) Segmentation Algorithms The search terms liver, vessel, 3D, ultras ound, reconstruction, snake, extraction, segmentation, and tracking were used to fi nd relevant literature. The most recent published review of vascular reconstruction techniques uncovered to date is Kirbas [42] in 2004 which is an exhaustive work th at categorizes reconstruction methods by approach. Twenty eight other more recent pa pers were reviewed and sixteen were found to be relevant to the missi on of autonomous vascular r econstruction from 3D ultrasound as there are many other segmentation algorithms that require the user to manually locate “seeds” [43] [44] [ 45] and otherwise manually direct the progress of the segmentation process. Segmentation approaches that re ly on pre-trained data or use anatomical libraries were reviewed to uncover ideas for fully automated approaches.

PAGE 55

34 Of all papers reviewed, only two and were specific to 3D ultrasound [46] [47], providing an apparent opportuni ty to advance vascular disc overy with this lower cost modality. Most of the reviewed articles employe d processes that benefit from venous contrast agents in conjunction with MRI a nd CT modalities. Only one paper treated zero contrast applications [48]. As contrast ag ents are not FDA approve d in the United States for ultrasound, the extraction of vasculature from Ultrasound images is more challenging [46] [47] [48]. As with real hepatic anatomical structur es, vessels captured in 3D images do not spatially overlap each other, eliminating the need to process for overlapping structures as in 2D processing [45], however, wave front propagation has been proposed as a means of detecting the centers of tubular structures which may be overlapping in 2D or close to each other in 3D [45]. References discuss the issues with scal ability of some approaches and offer “multi-scale” alternatives [45, 49-52]. Scaling problems occur when a filter must be optimized for the size of a structure and give s less conclusive results with larger or smaller versions of the structure. All papers indicate that the first processing step is to reduce the random noise in the image since the accuracy and decision making of all tracking algorithms will be compromised by voxels which are not representa tive of the tissue at that location. This noise specifically allows region-growing al gorithms to “bleed” beyond vessel walls. Noise reduction techniques included Mean-Gaussian [44] and Diffusion filtering[53].

PAGE 56

35 Intensity thresholds were used by some pr ocesses to remove non-vessel voxel “noise” from image volumes [49] Ultrasound data contains “speckle noise” [47] [27], a natural artifact of a coherent acoustic wave reflecting from objects and impe dance discontinuities. This noise can be either treated as a noise to be eliminated through filtering, or, may be processed to ‘provide additional information such as the density and number of random scatters based on the statistics of the received envelope’ [46]. This latter treatment of speckle is ‘useful in the discriminati on of breast cancer’. Discovery of hepatic vessel interiors can be performed by growing a manually or automatically placed seed volume into a larger volume which is constrained at the vessel boundary by voxel intensity chan ges or other rules based on force, entropy, energy or advection [43] [51]. Many previous investigators use a Hessian matrix as a filter to determine how similar an arbitrarily oriented structure is to a vessel [48] [4 5] [49] [52] [5 0] [54] [55]. The Hessian matrix is trained to the characte ristic values of a tube structure and the eigenvalues of the structure ar e used to provide an indica tion of “vessel-ness”. One process implements multiple eigenvalue sets to accommodate the identification of both tube-like and junction-like structure [54]. Hessian filters deal well with irregularly shaped voxel dimensions (anisotropy). When a 3D image is taken as slices, the vessel cross-sections appear as individual circles and ellipses. The centers in each slice are determined by graphical or mass centroid or “core” methods [56]. Vessel tracking may be accomplished by treating the

PAGE 57

36 movement of the vessel center as target tracki ng. This class of problem can be treated with Kalman filtering, a process developed for predicting the trajector ies of space objects and military targets [45] [47]. The Kalman f ilter process uses the present and previous states (center locations in each slice) of each circular ob ject to predict the center location in the next slice. As each new slice is processed and the new centers can be known, the weighting factors of the filter are adjusted to more accurately predict the next center location. The Kalman filter is very accurate for gradual tr ansitions and less accurate in sharp or reversing bends, a situation common in hepatic vasculature [47]. A number of papers present reconstruc tion by use of the intensity based voxel by voxel wise area and volume search. B-splines have been evaluated for establishing hepatic vessel centerlines as they establish a minimum curvature fit [53]. A geometrical moment method using a cylindrical shape as a vessel reference for reconstruction has been evaluated [56]. The tabulation of the approaches provide d in Table I.5 provides a summary of approaches to vascular reconstruction, however the documentation of success with ultrasound as a ‘noisy’ and ‘low contra st imaging modality’ was sparse.

PAGE 58

37 Table I.5 Summary of Vascul ar Segmentation Approaches Ref Modality User Input Noise Reduction Approach Vessel Segmentation / Detection Centerline Construction Vessel Surface Rendering [43] CT Seed and Contour Points Yes, not specified Region Growing N/A Skeletonization Surface of Region [44] CT Editing MeanGaussian Graph-cuts Centerline extraction Tree separation Medialness filter Radius from Medialness filter [45] 2D (3D) Xray No Not discussed Minimum cost Hessian Wavefront 3D not treated [48] CT Gaussian Median No contrast Hessian Thresholding Hessian [47] US Seed Kalman Median filter [27] US -Peak/Valley -Hilbert Curve -Cubic Spline -Windows adaptive threshold Otsu Core area Morphological filter [53] MRA 3D Diffusion Filter Voxel search 2nd Order Bspline [49] CT Threshold Hessian -Hough transform -Morphology [51] CT Edits Threshold Region Growing Hessian Hessian [52] MRA Hessian Hessian [50] MR Hessian Difference Hessian [54] CT Correlation Multiple Hessian Morphology Fuzzy shape [55] MRA Gaussian Multiscale Hessian [56] MR/CT Seeds Geometrical Moment BoykovÂ’s Graph-cut Min cut Max flow Preliminary Data Segmentation Applications Existing techniques for vessel segmenta tion and surfacing were explored using available software applications such as Geomagic and MeshLab as documented in the

PAGE 59

38 Portal Vein Flow Phantom fabrication (Cha pter IV) and the following discussion of an ITK-SNAP advection approach. Advection is a method of segmentation where a boundary between tissues or fluids is defined by the user as a specific ch ange or rate of change in voxel intensity. A “balloon-like” 3D manifold surface is then “inf lated/expanded” from a user selected seed point until the expanding surface reaches a bounda ry that meets the threshold or gradient rule and then stops at that voxel location. When all of the vo xels of the expanding surface reach a boundary, the final expanded su rface represents the geometry of the boundary surface and has an identical volume and surface area to the boundary surface. The Advection capabilities of ITK-SNAP (www.itksnap.org) were evaluated on several arterial and venous phase CT imag e volumes taken from liver donor studies. Liver donor studies were used since the succe ss of a transplant relies on good structural views of the donated organ in surgical planni ng. CT imagery was used for the evaluation since it is a much less “noisy” and higher resolution imaging modality than ultrasound, providing a benchmark of how well advect ion would perform on a good image. The reasoning was that if advection yielded marg inal segmentation on CT imagery, it would not be acceptable for ultrasound. As a contrast agent is used in the liver donor image acqui sition, the contrast in the arterial phase provides a distinct vessel wall as sh own in Fig. II.25.

PAGE 60

39 Figure I.25 Arterial Phase CT An advection segmentation of the desce nding Aorta and hepatic artery during the Arterial phase shown in Fig. II.26, where th e hepatic and splenic ar tery surfaces are well defined and reasonably smooth. Figure I.26 Arterial Phase Hepatic Artery In the venous phase, the contrast agen t has been diluted throughout the blood stream and filtered by the liver hepatocytes, resulting in lower image contrast and as shown in Fig. II.27.

PAGE 61

40 Figure I.27 Venous Phase CT The advection of this low contrast im age volume provided a much less accurate and unusable vessel wall definition (Fig II.28) due to the advection surface “leaking” into noise artifacts along the vessel wall. Figure I.28 Portal Vein

PAGE 62

41 The advection process is unable to disti nguish a legitimate voxel intensity change at a physical boundary from the same intens ity changes between voxels due to noise. Even using an established and well de veloped tool like ITK-SNAP on a fast laptop, the advection process took tens of minutes to complete which is not in line with the required TTR. The advec tion approach was not pursued further due since ultrasound imagery is more noisy than CT and the clinical time requirement was not met. Clinical Requirements Software Application A graphics based software applicat ion was the natural vehicle for the development of the algorithms required to estimate flow from Color Doppler and to capture their value in a clinically useful manne r. As such, there were clinician requested features directing usability and response tim e of the software application. The first directive was that the clinical application im plement a true 3D display that would allow vascular images to be viewed from any angl e or perspective. The second directive was that all functions other than the capture of a Region of Interest (ROI) around a tumor would operate autonomously without user dire ction or intervention. The third directive was that the clinical appli cation adhere to a Time-to-Re sponse (TTR) of 10 seconds to manage the cost and timeliness of the RF ablation pr ocedure. The TTR directive reflects that: clinician time is costly and a precious commodity; Operating Room rates can reach

PAGE 63

42 thousands of dollars per hour; and that the procedure is pe rformed on a conscious patient who would like to leave as soon as possible. Review of Existing Tools Prior to the development of custom so ftware, existing tools were reviewed for compliance with the research and clinical go als. The review included Osirix (Mac), ImageJ, IrfanView, 3D Slicer, Syngo FastVi ew, Onis, ORS, Qlab (Philips), ANSYS, SolidWorks, Meshlab, and Geomagic. Ultimatel y, the Philips QLab software application was the only available application capable of opening the 3D files from the available iU22 ultrasound cart. While QLab did pr ovide the desired 3D viewing and image manipulation, and manual and 2D automated segmentation, the application did not offer vessel center-lining, automated angle correction or flow estimations. Software Architecture MATLAB MATLAB was selected for initial algor ithm development and prototyping. MATLAB was not used in the development a nd deployment of the clinical algorithms and application due to the TTR requireme nt and Intellectual property concerns. MATLAB is based on JAVA, an “interpreted ” language that requires each line of MATLAB code to be broken down into JAVA code and then run on the JAVA Virtual Machine (JVM) every time a program is run. MATLAB documentation suggests that ‘time critical applications be developed in a “compiled” language such as “C” where the

PAGE 64

43 conversion to machine executable code happe ns only once. While MATLAB offers a C cross-compiler, the underlying lib rary algorithms are not nece ssarily converted into the most efficient constructs of the C language and, the code-compiled library functions carry an IP burden. As a collection of pre-developed libraries, MATLAB retains the rights to them through an expensive commerc ial license, a cost and burden that the project was not willing to bear. While MA TLAB does allow applications to be deployed using their client at no cost, it was deemed not feasible to add this client software to the laptops used in the hospi tal IT environment. The C Programming Language Other than being license free and widely supported, the most attractive feature of C is that it allows the direct and efficient manipulation of data in memory through the use of “pointers” and array subscrip ts, critical to fast processing of large images in real time clinical settings. Memory access in “obj ect oriented” languages such as “C++” or “JAVA” is performed through special class func tions, which slow the process. For nonmemory functions, C++ has features which are very useful for Gra phical User Interface (GUI) implementations and general control of applications and it integrates well with native C code. For the reasons stated above, the clinical application was implemented using the C programing language. To keep the clin ical application development process manageable, the process was broken into two steps: 1) implement the vascular

PAGE 65

44 reconstruction algorithms and basic graphics functionality in C and OpenGL; 2) migrate the application to the C++ based Qt user interface. Since the “C” programming language does not define or include a library of functions for the management of graphical data or user interfaces, the OpenGL (www.opengl.org) graphics librar y was selected. OpenGL is a free to use and distribute library when dynamically linked by the applica tion (static linking requ ires a distribution license), it is supported by all major operati ng systems, and is deployed on thousands of graphics systems and mobile devices. To provide the GUI and mouse based user controls and measurement reporting capability expected of a clini cal application, the C/OpenGL code base was imported to a user interface development environment called Qt. Qt was developed by Nokia for deploying interactive gr aphical applications across many stationary and mobile devices (laptops, tablets, cell phones) and ope rating systems (Windows, Linux, Android, MacOS). Qt was chosen to allow the clini cal application to be deployed to laptops, tablets and phones with “relative” ease and c onsistency. Google Earth is an example of a successful cross-platform Qt application. Th ere are no Qt licensing fees if the library functions are linked “dynamically” (i.e. D LLs), however, a distribution license is required if a static library is to be embedded within the application. TTR Compliant Architecture As Operating Room and clinician over head costs are high, the algorithms and applications were designed to minimi ze processing and calculation times. The

PAGE 66

45 application architecture is based on fixed size pre-allocated static integer arrays for one time storage of the raw image data. This eliminates the time and need to re-scan the image data from files and the time required to dynamically allocate and de-allocate the very large arrays which are up to 1720 x 1024 x 180 x 3 Bytes (1TB) for the Visual Human Images and 512 x 512 x 512 x 3 Bytes (500 MB) for 3D ultrasound data. A side benefit of static arrays is that they can be successfully implemented under any Operating System (OS) (a clinical deployment goal). Many Operating Systems do not have robust allocation/de-allocation functions, which can result in “memory leakage” and system crashes that make the application appear uns table and unusable. The benefit of using a static integer array is that it is smaller th an using a floating poin t data type. As the volume of vessel structure is a very small percentage of the total image volume, the vessel cross-sections and propert ies are also stored as “obj ects” in a “list” (a single dimension linear array) as they are discovered. With raw data available in the static 3D array, functions that involve relative and abso lute indexing to voxels of interest can be performed very quickly with integer math, a nd, vessel objects, features and parameters can be quickly picked from a list. This combined approach enables very quick calculations and the best possi ble user response times.

PAGE 67

46 Development Environment To facilitate speed of development, a Dell Precision M4500 laptop PC was used with the following configurat ion: Operating System Wi ndows 7 64 bit, CPU Quad Intel I7 2.4Ghz, RAM 8GB, Graphics NVIDIA Quadro FX 1800M; Programming Environment MATLAB, Qt, DevC++; C AD Applications MeshLab, Geomagic, SolidWorks, ANSYS.

PAGE 68

47 CHAPTER II RESEARCH Specific Aims The development of autonomous flow estimation from 3D Doppler ultrasound was broken into steps to guide the developmen t and verification of i ndividual algorithms. Aim 1 – Fabricate and Model a Portal Vein Flow Phantom In order to validate hepatic flow estimate s it was necessary to create anatomical and geometric flow phantoms, and perform flow modeling. Aim 2 Obtain an ROI from a 3D Image Volume from Clinician Input To begin the flow estimation process, it was necessary to acquire 3D image data from an ultrasound cart and provi de the clinician with a gr aphical means of specifying what Region-of-Interest (RO I) of the 3D volume contained the tumor of interest. Estimating the flow within a sub-volume of the entire 3D image was deemed a more achievable and more TTR compliant task. Aim 3 Autonomously Segment Va sculature and Compute Centerline With an ROI selected, it was necessary to find and identify all of the unique vessel segments within the ROI accommodating ve ssels that may be very close together

PAGE 69

48 and be in proximity to spatial noise artifacts. It was also necessary to find each vessel segmentÂ’s centerline to facilitate diameter determination and angle correction of velocity. Aim 4 Autonomously Compute Segmente d Vessel Diameters and Cross-sectional Areas The next step in flow estimation invol ved finding the vessel diameter from its cross-section (which is normal to the centerlin e direction) and compu ting the area of that cross-section. Aim 5 Autonomously Compute Angle Corrected Velocities for Segmented Vessels Because Color Doppler measures velocity relative to the beam direction, it was necessary to correct th e Color Doppler velocity data in the direction of the vessel at the point of intersection on the vesse l centerline to obtain the velo city in the direction of the vessel. This required a determination of the transducer origin for angle correction with the centerline estimated in Aim 3. Aim 6 Autonomously Compute Volumetric Flow for Segmented Vessels Estimated flow was obtained by multiplying the angle corrected velocity and cross-sectional area at the points of interest along the centerline Each algorithm developed in the Aims above was evaluated using flow phantoms of known physical geometry a nd calibrated flow rates.

PAGE 70

49 Materials and Methods Aim 1 – Fabricate and Model a Portal Vein Flow Phantom Portal Vein Phantom As a study of hepatic vascular flow, indus trial casting materials were researched and imaged with ultrasound, MRI and CT, and th en used to fabricate flow phantoms. Materials Study A set of usability and performance crit eria were established for a proposed phantom design capable of supporting liquid fl ow during imaging. A literature search was conducted to identify the materials and methods previously used in phantom fabrication. A database of human tissue and casting material properties was compiled to facilitate the selection of appropriate mate rials for testing. Several industrial casting materials were selected, procur ed, and used to fabricate test samples that were imaged with Ultrasound, MRI and CT. The full text of the published phantom materials study [57] is provided in Chapter III. Fabrication A portal vein structure was extracted from the Visual Human Male data set for use as a pattern in the flow phantom construction. A three-dimensional (3D) portal vein pattern was created from the Vi sual Human database. The port al vein pattern was used to fabricate two flow phantoms by different methods with identical interior surface geometry using Computer Aided Design (CAD) software tools a nd rapid prototyping techniques. One portal flow phantom was fabr icated within a solid block of clear P-4

PAGE 71

50 silicone for use on a table with Ultrasound or within medical imaging systems such as MRI, CT, PET or SPECT. The other portal flow phantom was fabricated as a thin walled tubular dipped latex structure for use in water tanks with Ultrasound imaging. Both phantoms were evaluated for usability and durability. The full text of the published phantom fabrication [58] is provided in Chapter IV Image Testing The image quality of the perfused solid si licone flow phantom was tested on a lab bench with the 3D transducer coupled to th e phantom body with acoustic gel as shown in Fig II.1. Figure II.1 Solid Silicone Body Portal Phantom and Test Setup

PAGE 72

51 The image quality of the perfused thin wa lled latex phantom was tested in a water tank with the transducer and phantom immersed as shown in Fig II.2. Figure II.2 Thin Walled Latex Po rtal Phantom and Test Setup Portal Phantom Flow Testing The branch outflows of the thin walled la tex portal phantom were measured using a mass/time method over a wide range of flow rates and pressures taken from Table I.1 and the test setup shown in Figs II.2-3. The adjustable height (0”32”) gravity tank used to adjust head pressure is shown on the left of Fig. II.3 and the individual branch outflow tubes are shown on the right.

PAGE 73

52 Figure II.3 Thin Walled Latex Portal Phantom Test Setup The water outflow of each phantom branch was collected over a fixed period of time and was then weighed to provide mass flow rates in kg/s, providing a mean mass flow rate. CFD Simulation A CFD flow model and simu lation was created using ANSYS CFX software. The portal vein flow phantom mesh documented in Chapter IV was imported as a pattern dependent mesh into the ANSYS ICEM mesh ed itor so that all surface inaccuracies in the portal mesh would influence the simulation results. The inlet boundary conditions were set to flow rates and pressures taken from the portal phantom testing as discussed above and shown in Figs. II.2-3. The branch outlet boundary conditions were set to a constant

PAGE 74

53 backpressure representing the 9” (22.9 cm) of head height in the portal vein branch outflows. Water was selected as the fluid and the defau lt ANSYS properties were used. Portal Vein Phantom and CFD Comparison The results from the measured phantom flows and the CFD simulation were converted to volume flow rates and plotte d together for comparison. The mean Mass flow rates from the branch measurements were converted to a mean Volume flow rate using Eq. II.1 and the density of water at 0.001 kg/cm3. Equation II.1 Mean Mass to Volume Flow Conversion Pulse Wave Doppler velocity measurements were not included in the comparison due to the lack of reliable physical regi stration between the CFD simulation and PW Doppler measurement points along the phantom vessels. Verification Phantom Setup The clinical version of the software a pplication (VAS) and developed algorithms were evaluated using simple flow phantom s comprising tubing of known diameter at varied flow rates and transdu cer beam sweeps. Water was us ed as a test fluid and dry corn starch was dissolved to increas e the echogenicity of the solution.

PAGE 75

54 Geometrical Tubing Phantoms were fabricat ed using latex tubing of varied sizes to enable testing of individual and grouped tubes as shown in Fig. II.4. Figure II.4 Thin Tube Phantoms The “parallel” phantom is show in the left image of Fig. II.4 where the 3/8” (0.925 cm), 1/4”(0.635 cm), and 1/8”(0.317 cm) latex tubing segments were connected to the inlet and outlet manifolds with independent valves cont rolling the flow through each tube. The parallel phantom was used for practicing the capture of non-aliased Color Doppler flows utilizing the full veloc ity scale of the ultrasound cart. The “series” phantom is shown in the mi ddle image of Fig. II.4 where the 3/8”, 1/4” and 1/8” latex tubing segments were connected end to end w ith the 3/8” accepting the inlet flow and the 1/8” tube carrying the outlet flow, with all flow controlled by an outlet valve. The series phantom was used to investigate the color range of the Color Doppler velocity scale.

PAGE 76

55 The “net flow” phantom is shown in the right image of Fig. II.4 and used a 3/8” latex tubing with a 1/16” wall in a serpentine la yout to allow the same flow to transit the image volume in both directions. The net flow phantom was used to test the clinical application’s ability to measure vessel diam eter, vessel area, angle corrected velocity, inlet and outlet flow rates and, and zero net flow through the ROI. Net Flow Phantom Calibration Two methods were used for measuring st eady flow through the net flow phantom and the results from several image samples were compared at several flow rates. Mass/Time The mass of water exiting the net flow phantom over 2 a minute interval was measured on a scale to provide a mean Mass fl ow rate. The time interval was measured with a stopwatch. The mean Mass flow rate was converted to a Volu me flow rate using Eq. II.1. Angle Corrected Velocity Using the ultrasound cart’s Pulse Wave (P W) Doppler Sample Volume and Angle Correction tool, the Angle Corrected Velocity of the fluid in the phantom tube was measured at several locations along the flow in cm/s. The Volume flow rate was obtained by using Eq. II.2 using the phantom tubing diameter of 3/8” (0.925 cm).

PAGE 77

56 Equation II.2 Velocity to Volume Flow Conversion The measurements from the ultrasound carts were: 1) Time Averaged Mean Velocity (TAMV) obtained by averaging all of the highl y variable velocity data; and 2) Time Averaged Peak Velocity (TAPV) obtaine d by averaging the maximum velocities (velocity peaks). For laminar and parabol ic profile flow (Poiseuille), the peak (centerline) velocity is expect ed to be 2x the mean value. For turbulent and flat profile flow (Plug), the peak and mean veloci ties are expected to be the same. Flow Estimation Verification To expose issues with the velocity es timation methods, five randomly selected ROIs were taken from images of the tubing in the net flow phantom which were acquired under different flow conditions. The images of the net flow phantom were taken with the transducer centerline at an a ngle of 45 to the net flow phantom tubing with the beam scanning +/30-40 from the transducer cen terline. To utilize the maximum dynamic range for each of the samples, the velocity sc ale and flow rates were adjusted such that maximum scale value was slightly higher than the highest Doppler velocity detected in the image, with a slight marg in to prevent aliasing artifact s. The velocity ranges were selected to test the performance of the au tonomous algorithms, which may create laminar and/or turbulent flow conditions. The Re ynolds numbers of the centerline and crosssectional mean velocities were both computed to help determine the nature of the flow.

PAGE 78

57 The following parameters were measured and recorded: 1) Maximum Velocity Scale value fr om the ultrasound cart color/velocity display 2) The Average of 3-5 manually angl e corrected Time Averaged Mean Velocity and Time Averaged Peak Velocity taken from sample volumes along the net flow phantom tube 3) The Cross-sectional Mean of the Angle Corrected Inlet/Outlet Velocities taken from all of the voxels captured in the cros s-sections at the top and bottom of the ROI 4) The Centerline Angle Corrected In let/Outlet Velocities taken from the top and bottom centerline voxels within the ROI Aim 2 Obtain an ROI from a 3D Image Volume from Clinician Input Obtaining 3D Image Data Access to a Philips (Bothell, WA) iU 22 Ultrasound system (SN 02L6HO, Rev. 6.0.0.X) and a V6-2 curvilinear 3D transduc er was obtained for the capture of 3D ultrasound images. Philips provided a precompiled (.p) MATLAB library function that allowed access to their proprietary 3D file format. Displaying 3D Image Data and Capturing an ROI Algorithm Development Platform (ADP) To facilitate the examination of ra w data from the ultrasound cart and development of prototype algorithms, a cust om GUI based viewer named the Algorithm Development Platform (ADP) was develope d in MATLAB. The ADP Graphical user interface was designed to allow: viewing image data in two viewing panes; viewing data from each of the lateral, depth, and elevati on views; data to be viewed in slices

PAGE 79

58 synchronized between the BW and Color D oppler data; the development of vessel finding, segmentation, and ROI selection functions. The ADP GUI is shown in Figs. II.5. Figure II.5 BW and Doppler Slices (BW Mode) The default mode of the ADP is View mode set by the View/Build button which allows the developer to view the raw imag e data or data created by the functions developed for vascular recons truction as described below. The ADP was designed to allow 2D slice by slice viewing of 3D data in all axes (Depth, Lateral, Elevation) of the dataset both Forward and Reverse order using the DF, DR, LF, LR, EF, ER buttons. All operations were started with the Run butt on, paused with the PSE button, and stopped completely with the Stop button. The rate of slice display was controlled by the Delay (DL) button and the slice count in the imag e was displayed in a text box. The Slice (SLC) button allowed stepping th rough the image volume on a slice by slice basis. The

PAGE 80

59 Test (TST) button forced the ADP to use pre-co nfigured test image volumes to assist with image registration and orientation. The color mapping of the image data is selected with the COL button and the color gamut was displa yed as a bar on the right margin of the GUI. As MATLAB (Windows OS) applies a co lor map on a global (not window) basis, both panes are forced to the same color ma p. Images mapped to a linear BW map are shown in Fig. II.5 and images mapped to a li near false color map are shown in Fig. II.6. False color mapping has no meaning for BW images. Figure II.6 Geometry of Doppler (False Color Mode) The ADP was designed with two viewing panes with independe nt and identical controls such that a BW ultrasound image sli ce could be displayed in one pane and the identical slices of the Color Doppler data co uld be shown together in the same pane as shown in Fig. II.6, or, the red Doppler data (flow towards the transducer) could be

PAGE 81

60 displayed in one pane and the blue Doppler data (flow away from the transducer) could be displayed in the other pane as shown in Fig. II.7. Figure II.7 Red and Blue Doppler (False Color) As utility functions, the HST button plots the histograms of color/velocity values within the selected image data as shown in Fig. II.8.

PAGE 82

61 Figure II.8 Histograms of Red and Blue Doppler The CMP button displays the activ e color map shown in Fig II.9. Figure II.9 Linear Color Mapping of BW and Doppler

PAGE 83

62 The ROI button creates orthogonal side vi ew and a movable selection box to allow selection of a 3D ROI. The ADP was implemented in 2500 single spaced lines of original code. Vascular Analysis System (VAS) (C/C++/OpenGL/Qt) [Ultrasound Data] The VAS user interface is shown in Fig. II.10. Figure II.10 “VAS” Clinical Appl ication Graphical User Interface Leveraging OpenGL capabilities, Ort hogonal, Perspective and Navigational viewing modes were implemented. The Orthographic and Perspective modes allow the clinician to rotate, spin and translate the image around a fixed focal point at the origin in

PAGE 84

63 the center of the vasculature image with cont rol over the eye-to-origin distance. In Orthographic mode, the data is shown with righ t angles. In Perspective mode the data is shown with a vanishing point as it would be seen in a photograph or painting. Navigation mode allows movement of the point of focu s in the scene and allows the clinician to “travel” towards and away from the focal poi nt as if the viewer was walking through the image volume. Navigation mode is also called “fly through” m ode and allows the clinician to see the inside of vessel surfaces for inspection of stenosis, branching and other vascular pathologies by viewing them form within the structure as a small observer. The Up/Down/Left/Right buttons pan the scen e in Perspective mode and shift the focal point in Navigation mode. The Axes button tu rns the white x, y, z reference axes on and off. The view of the object can be set to an orthogonal view at any time with the Depth/Lateral/Elevation directi on from the back or front. The default Orthogonal view can be recovered with the Reset. The Sli ce button enables the clinician to enter an arbitrary slice, or, increment/decrement by one slice to step through the image. Buttons at the bottom of the display en able toggling independent disp lay of the Black and White, Red Doppler, Blue Doppler data. Since OpenGL plots each voxel as an indi vidual point and th ere are typically 368 x 272 x 256 voxels = 26 Million voxels in the BW image alone, the screen update time for user requests to rotate or zoom the view can take several seconds, making quick inspections of data difficult. To accomm odate faster viewing, the application was designed to operate in Cloud mode, where only the voxels representing the outside surfaces of the image volume are plotted to provide a smooth and real time response to

PAGE 85

64 the user. The surface clouds of the BW, Red Doppler and Blue Doppler are shown in Fig. II.11. Figure II.11 White BW Scan Envelope Overlaid with Red and Blue Doppler A further benefit to cloud mode is the ability “see through or past” neighboring voxels for a more intuitive view of the 3D data geometry, particularly for structures that intertwine or obscure each other. To enable the user to view all of the data points in the selected image volume a raw data viewing mode can be invoked with the RAW button. To provide some level of transparency in raw data mode, the points plot ted for each voxel are slightly smaller than the voxel volume to allow the user to see though and past some number of surface voxels to gain insights into the texture of the interior data as can be seen in Fig. II.12 for the BW

PAGE 86

65 volume (where the compound curvature of the 3D transducer can be seen at the top of the image), and in Fig. II.13 for the Red and Blue Doppler volumes. Figure II.12 Raw BW Data Volume (Semi-transparent) Figure II.13 Red and Blue Doppler Data in Raw Mode

PAGE 87

66 In Cloud mode, the voxels are plotted as saturated (color on/off) “false” RGB colors for visibility. In Raw mode, the voxels are plotted in linearly mapped (min voxel data value = no color, max voxel data value = max color) “false” RGB colors relating low voxel value to low saturation hue and high voxel value to high saturation hue as a means of providing the user with relative velocity information and better visibility to low intensity voxels. RawTC mode presents the image data as it would be seen on the iU22 LCD using the Philips velocity/color map. LCD displays implement an “additive” color scheme where all colors “off” displays the only the white from the florescent or electroluminescent backlight (white light pass ed), and all colors “ on” deepens their hues to present a black display (w hite light blocked). The Ra wTC color mapping transitions from black surrounding the vessel to a dark hue of the color represen ting low velocity at the vessel wall and then through lighter hues of the color to near white for the highest velocity at the center of vesse ls. The hue/velocity variati on in the vessel cross-section shown in Fig II.14 suggests a Poiseuille (p arabolic) velocity profile, where “Plug” (uniform) velocity profile would be indicat ed by a uniform color hue across the vessel cross-section.

PAGE 88

67 Figure II.14 Red and Blue Doppler shown in RawTC (True Color) Mode Importing Data Ultrasound images captured on the iU22 ar e transferred to th e Laptop PC running the VAS development environment using a USB memory stick. The image file is opened by a conversion program developed in MATLAB which opens the file using the Philips (*.p) library function and then e xports the data as four separa te files: three binary image arrays (BW, RED, BLUE), and a text (*.t xt) file which contains the x, y, z image dimensions and floating point voxel dimensi ons. These files are imported by VAS for use by the developer and clinician.

PAGE 89

68 Capturing a Region of Interest (ROI) After the clinician selects an image to import, VAS scans each data file and creates the surface point cloud of the Doppler data after whic h it is available for display and ROI selection in the main window. The ROI Set, ROI Clear, and ROI View f unctions allow the clinician to manually select the ROI. The clinician can set the bounding box using a Size and Position method, or, by setting the Upper Left, Lower Right, Le ft and Right sides of the ROI box, or, as a future enhancement, by Dragging the box size and position. The 3D corners of the box are determined by the inters ection of boxes projected from two orthogonal views. The clinician first sets the ROI box size and position in the late ral view, the view is then switched to Elevation view wh ere the box is sized and posi tioned to complete the 3D corner selection. A click of the right mouse button steps the clinician through each selection, or, the GUI buttons can be used to direct the selections and completion. VAS was implemented in 7500 single sp aced lines of original code. Aim 3 Autonomously Segment Va sculature and Compute Centerline Segmentation The initial method developed for finding the external surfaces of the Doppler data was a raster scan and threshol d edge detection algorithm that was implemented within the function responsible for the in itial reading of the Doppler data file for processing efficiency. This method was used in both ADP and VAS.

PAGE 90

69 ADP Segmentation To capture circular and slightly elliptical cross-sections of sliced vessels and not vessels cut along their length, a Masking (spa tial filtering) function was developed and invoked by the MSK button on the ADP GUI. As ultrasound image data contains spatial noise (volumetric blobs unrelated to vessels ) which would interfere with the vascular surface reconstruction, the MSK process wa s assisted by the Noise Reduction (NR) function which was developed to eliminate cr oss-sectional objects with an area below a size set by the slider control labeled NR us ing the MATLAB “bwareaopen” and “imfill” library functions. After an image is pr ocessed by MSK and NR, a Contour (CNT) function identifies the perimeter of the mask id entified objects which ar e plotted as a blue overlay on the cross-section; a Center (CEN) function co mputes the object’s center overlaid as a red circle; and an elliptical re presentation of the object is overlaid as a red ellipse. These functions utilize the MATLAB “contour”, “bwconncomp” and “regionprops” library functions. Smoothing of the Color Doppler surface was implemented in a function invoked by the (SUR) which utilizes the MATLAB “smooth3” library function and a 9x9x9 voxel Gaussian convolution kernel. The “sm ooth3” function convolves each voxel of the Color Doppler data volume with the kernel th at has a Gaussian weighting of one at the center voxel and tapers the values of the surrounding voxels in the cube to smaller numerical values at the edge of the kernel cube following a Gaussian distribution. This process effectively shifts voxels from their original positions to nearby positions to create a smoother spatial surface.

PAGE 91

70 VAS (C/OpenGL) Segmentation [Visual Human Data] TolTech (Aurora, CO) provided an STL/OBJ file format mesh for the portal vein structure that was used to fabricate the flow phantoms described in Chapter IV [58]. The mesh provided was created using matrix methods that applied uniform threshold rules to the entire image data set, and as such, was not able to re-create the mo re subtle details of local vessel geometry. The provided mesh also contained many fragments and holes which required considerable manual smoothi ng as described in Chapter IV [58]. However, the database for the provided mesh was also the basis of TolTech’s commercial anatomy training software VH Dissector Pro which provides high resolution [1760 x 1024] head-to-foot photographic images of slices of a human male with manually segmented hepatic vasculature used for medical training. For improved hepatic modeling, the desired hepatic vascular struct ures were extracted from these images by activating the segmentation detail overlay fo r the portal vein, hepatic vein, and liver capsule individually, and then ma nually saving each of the 180 superior to inferior axial slices comprising the liver volume as individual BMP image files. As the manual vessel segmentations app eared as purple/turquoise color overlays in the images, Adobe Photoshop was used to identify the color gamut values of the component Red/Green/Blue (RGB) hues su ch that the segmentation could be discriminated from the general yellow/brown anatomy. The “C/OpenGL” version of VAS was wr itten to open each image file, capture the dimensions and shape of each vessel crosssection, and then display in 3D “virtual reality” the resulting point cloud for the user to view. The serial data stream for each

PAGE 92

71 image was read from the BMP file as a horizontal then vertical image raster scan of voxels (pixels in the 2D plane). The inco ming voxel values were compared to the RGB color range of the Visual Human segmentation color. For voxel values within the RGB color criteria, the leading and trailing voxels positions of the vessel cross-section were recorded and then the halfway point between them was computed and recorded. This process created a list of line segments wh ere the endpoints represen ted the voxels of the surface point cloud. Line segment centers were used to diffe rentiate vessel cross-sections in close proximity since line segment centers are more central to the vessel cross-section and thus further away from each other than the edge voxels of two adjacent ve ssel cross-sections. Vessel Edge Detect Pseudo-code for all Visual Human BMP files (n=180) for all horizontal raster lines in the BMP file for all voxels in the current raster line for each line segment meeting segmentation color rules record the first voxel position record the last voxel position compute the difference between the fir st and last voxel as the segment center Vessel Object Pseudo-code for each slice in the image volume (n=180) for all the line segments in the slice find the line segments centers that are within q voxels of each other and assign them to a unique vesse l cross-section object

PAGE 93

72 The center of vessel cross-sections were calculated using an average center of mass method. The gross dimensions of the vessel cross-se ctions were detected and saved. The segmentation function identifies the edges of the vessel (ends of lin e segments) with red dots, line segment centers with green dots, the mass centroid as a single white dot, and the maximum vessel dimensions with a grey bounding box. Further C functions were written using th e OpenGL graphics language to create a viewing window and keyboard based controls so that the clinician can view and manipulate the point cloud of th e vascular structure as a 3D object in a “virtual world”. The portal vein structure is displayed in re d, the hepatic vein in green, and the liver capsule in blue. As the clinic ian changes the point of view, the integer voxel positions in the static array holding the image data are tran slated into floating point values using the physical voxel sizes provided by the ultrasound cart and are th en plotted in the OpenGL viewing volume. Vessel Center Pseudo-code for all slices in the image volume (n=180) for all vessel cross-section objects for all the line segments in the object sum the x values of the segment centers sum the y values of the segment centers Divide the final x and y sums by the number of segment centers to get the average x and y location of the mass center

PAGE 94

73 Total Volumes A function was created to sum the volume of the voxels captured with within the portal vein, hepatic veins and liver capsule using the floa ting point voxel dimensions. CAD File Export A function was created to export the voxels representing the portal vein, hepatic vein, and liver capsule surface point clouds in a standard (*.xyz) file format for importing to other Computer Aided Design (CAD) software. VAS (C/C++/OpenGL/Qt) [Ultrasound data] In the “C/C++/OpenGL/Qt” version of VAS, a “nearest neighbor/infection method” of identifying and tagging vessel segm ents was developed. Each vessel segment identified was color coded for ease of viewi ng. Each voxel captured in the ROI is tested for association to an object, where voxels near voxels already associated with an object are assigned to the same object, and voxels not near assigned (i nfected) voxels are assigned a new object number. The ROI volume is scanned systematically to ensure all voxels are assigned and color coded by object number for identification when viewed. The spatial capture range (infection distance) can be adjusted to accommodate varied types of noise and close proximity of the vesse ls. Use of small capture ranges is good for Volume Pseudo-code for all the slices in the image volume (n=180) for all the line segments in the slice compute the number of voxels betw een the line segment end points sum the voxel count multiply the sum by the voxel dimensions to get total volume

PAGE 95

74 discriminating close objects at the risk of an object being fragmented into multiple objects if it contains intern al voids which are larger than the capture range. Centerline ADP Centerline As part of the Contour (CNT) f unction, the MATLAB “regionprops” library function was called to identify the cente r of vessel cross-sections found by the segmentation method. VAS Centerline A special centerline finding algorith m was developed using a mass centroid approach to create line segments whic h approximate the centerline of each vessel captured within the ROI. The algorithm partit ions the ROI into an adjustable number of compartments defined by planes perpendicu lar to the beam direction. A mass centroid algorithm is used to identify the center of mass of the compartment. Vessel Segmentation Pseudo-code for all voxels in the ROI test for association to an object within an adjustable capture range if voxel not already assigned if there is a voxel assign ed within the capture range assign the object number else assign a new object number

PAGE 96

75 Line segments are drawn graphically between the compartment centers to create the user visible centerline. Aim 4 Autonomously Compute Segmente d Vessel Diameters and Cross-sectional Areas Diameter ADP Diameter A function was written to calculate the vesse l cross-sectional diameter using the MATLAB image processing f unction “regionprops”. VAS Diameter Equation II.3 shows that any voxel B forming a vector CB with voxel C on the centerline which satisfies the Do t Product rule of AC • CB = 0 with the vessel centerline AC within a set tolerance, lies on the crosssection of the vessel and defines a vessel radius of CB which is half the vessel diameter. Centerline Pseudo-code divide the ROI into n compartments for all compartments for all voxels with in the compartment sum the x values of the voxels in the compartment sum the y values of th e voxels in the compartment sum the z values of th e voxels in the compartment Divide the final sums by the number of voxels in the compartment to get the x,y, z locations of the mass center

PAGE 97

76 Equation II.3 Voxels in the Vessel Cross-sectional Plane Once the on-plane crosssectional voxels are loca ted, the distance from the centerline to each voxel is located and the furt hest three voxels were averaged to provide a radius for use in the vessel cross sectional area calculation de scribed below. The radius averaging accommodates a small amount of eccentricity in the cross-section. The estimated diameters were compared to the phantom tubing inner diameter to calculate the percentage error as (Dest – Dr ef)/Dref so that la rger than reference diameters would give a positive error. Diameter Pseudo-code for two centerline segmen t points on the centerline (top and bottom of ROI) for all voxels within the ROI if voxel satisfies do t product within to lerance of +/-0.0001 tag voxel as being on cross-section compute distance from centerlin e to voxel and store the 3 longest average the longest three rad ii and double to get diameter A B C

PAGE 98

77 Area ADP Area A function was written to calculate the vessel cross-sectional area using the MATLAB “regionprops” library function. VAS Area For speed and ease of calculation, the geom etric form for the area of a circle based on radius (Eq. II.4) was implemented. Equation II.4 Area of Vessel Cross-section Aim 5 Autonomously Compute Angle Corrected Velocities for Segmented Vessels Velocity ADP Velocity Velocity algorithms were not develope d on the ADP as all development migrated to VAS. VAS Velocity The velocity of the fluid captured within each voxel is provided as a value in the decimal range of 0-255. Since there may be many voxels (points of unique velocity) in the cross-sectional capture, the velocities of all of the voxe ls captured are averaged to

PAGE 99

78 create a mean for the cross-section as shown in Eq. II.5 where n is the number of voxels in the cross-section. Equation II.5 Cross-se ctional Mean Velocity Use of cross-sectional mean velocity is important when Poiseuille (parabolic) flow is present since the veloci ty at the center of the flow may be much greater than the flow at the walls of the vessel where the velo city is zero. As discussed in Chapter I, Poiseuille flow is found downstream of br anching points and bends where the fully developed flow takes on a parabo lic velocity profile. The us e of a cross-sectional mean velocity calculation for a plug flow profile is not an issue as it is already a constant. Centerline velocity values were obtained from individual voxels along the centerline. VAS Angle Corrected Velocity As discussed previously, computing accurate flow in the direction of the vessel centerline requires an an gle correction of the Doppler velo city data. To automate the process of angle correcting vessels captured in the ROI along the cen terlines of captured vessels, an algorithm was developed to auto matically determine the focal points of the transducer fan beam by detecting the edges of the BW data cloud, computing a line fit along the edge lengths, and displaying the poi nts of intersection. Spatial filtering was implemented to detect the noisy BW fan beam envelope edges accurately. The BW

PAGE 100

79 volume edges are located by sampling the BW data volume for non-zero voxels at five levels in the depth axis by propagating a plane of 40 x 40 voxels halfway into the BW data volume in the lateral dir ection. If no voxels are touche d by the plane, the plane is stepped in the elevation direc tion and the process repeated un til voxels are contacted. As soon as an edge is detected the algorithm m oves on to the next sample point in depth. A line is then drawn through the points of edge contact and proj ected towards the transducer and is mirrored to the othe r three corners in two steps to provide intersections representing the lateral (phase d array) sweep origin and the elevation (mechanical) sweep origin. The 2D projection of th e lateral sweep origin is used as the transducer origin as that is the origin used for the 2D Power D oppler angle correction on the ultrasound cart. For autonomous angle correct ed velocity, the two end segments (ROI inlet and outlet) of the computed centerline are use d. The Cosine of the angle between the insonation beam from the transducer origin to one end of each centerline segment is computed using a trigonometric form of the Dot Product relationships (Eq. II.6). A = vessel direction B = insonation direction = angle between insona tion and vessel direction Vdop = voxel-wise Color Doppler velocity Vac = voxel-wise angle corrected velocity Equation II.6 Angle Between Insonation Beam and Centerline Segment

PAGE 101

80 The velocity estimates were compared to the mass flow and values measured on the cart to obtain a percentage error as (Vest – Vref)/Vref so that la rger than reference velocities would give a positive error. The Time Aver aged Peak Velocity (average of maximum peaks) from the ultrasound cart measurements was used as Vref for the centerline error calculations. The Time Averaged Mean Velo city from the ultras ound cart measurements was used as Vref for the cross-section mean error calculations. Aim 6 Autonomously Compute Volumetric Flow for Segmented Vessels The initial method for computing inlet and outlet flows of each vessel segment captured within the clinician set ROI uses the area and angle corrected flows obtained from previous aims as inputs to Eq. I.2. The net flow through single and multiple tubes was captured from multiple images and computed as the percentage diffe rence between the inlet (centerline segment 0) and outlet (centerline segment 1) volume flow values divided by the average of the inlet and outlet flow rates as (F seg0 Fseg1)/(Fseg0 + Fseg1)/2.

PAGE 102

81 Results and Discussion Aim 1 – Fabricate and Model a Portal Vein Flow Phantom Phantom Materials Selecti on Results and Discussion Five silicones and one polyur ethane were selected for testing and samples of all materials were successfully fabri cated as shown in Fig. II.15. Figure II.15 Phantom Test Samples All imaging modalities were able to discri minate between the materials tested and water references. Ultrasound testing showed th at three of the silico nes could be imaged to a depth of at least 2.5 cm (1 in). The RP-6400 polyurethane exhibited excellent contrast and edge detail for MRI phantoms and appears to be an excellent water reference for CT applications. The 10T and 27T silicone s appear to be usable water references for MRI imaging. The full text of the published phantom materials study [57] is provided in Chapter III.

PAGE 103

82 Portal Phantom Fabricati on Results and Discussion Solid silicone body and thin latex phantoms were fabricated successfully using rapid prototyping processes and passed durab ility criteria. The full published text detailing their construction [ 58] is provided in Chapter IV. The solid body phantom is shown as a void in the shape of the portal vein within the clear P-4 silicone as shown in Fig. II.16. Figure II.16 Solid Silicone Body Portal Vein Phantom The slight cloud in the center is residual hum idity from the pattern dissolving process. The opacity on the sides of the phantom is fr om the stepping resolution of the 3D printer used to fabricate the mold body. The thin walled latex phantom is shown in Fig. II.17 with the same portal vein shaped internal void as the solid body phantom.

PAGE 104

83 Figure II.17 Thin Walled Dipped Latex Portal Vein Phantom Phantom Imaging Results and Discussion Solid Body Phantom The goal of capturing the entire phantom volume in one 3D sweep was not met due to excessive attenuation in the P-4 material. Vessels de eper than 1” (2.54 cm) could not be imaged. While the P-4 showed good imag e quality at 1” (2.54 cm) in the materials study test sample, this turned out to be an in sufficient test for larg er scale castings. In an attempt to obtain portal structure images, a significant amount of material was removed from the solid silicone phantom to reduce the thickness to less than 1” (2.54 cm). The resulting BW and Color Doppler ultrasound image captu re of a vessel and outlet port is show n in Fig. II.18.

PAGE 105

84 Figure II.18 Solid Body Silicone Portal Vein Phantom The BW lumen of the surface vessel in Fig. II.18 was fuzzy and indistinct, and the Color Doppler flow is not visible along the length of the vessels. The solid body phantom was not used for further ultrasound testing; however it will be evaluated in future MRI and CT testing.

PAGE 106

85 Thin Walled Phantom A Color Doppler image of the thin wa lled phantom is shown in Fig. II.19. Figure II.19 Thin Walled Latex Portal Vein Phantom The thin walled latex phantom provide d excellent lumen detail with Color Doppler. Pulse Wave Doppler velocity m easurements were easily acquired through the lumen walls.

PAGE 107

86 Thin Walled Phantom Flow Testing The flow and pressure relationships for the portal vein inlet are shown in Fig. II.20 with boxes representing the ra nge of normal and peak portal values from Table II.1. Figure II.20 Portal Inlet Flow Vs Pressures The portal inlet velocities shown repres ent the upper operating limit of the portal vein flow phantom. In future work, the addi tion of resistance to flow in the branch outflows will simulate hepatic tissue a nd reduce the flow rates to normal ranges. The absolute and relative outlet branch flow rates as a function of portal inlet pressure are shown in Fig. II.21. The bran ches were numbered and matched to the name of their respective major supply vein (Left Portal Vein, Right Anterior Portal Vein, Right Posterior Portal Vein) and their respective regions (I-VIII) of the liver perfused by those veins and branches as described in Fig. I.12.

PAGE 108

87 Figure II.21 Portal Branch Flow Rates The branches of the RAPV handle the majority of the flow which is understandable as the right lobe of the liver has considerably more mass that the left lobe. The ratios of the branch velocities and flows are relatively co nstant across the pressure ranges with some re-distribution of flow at the normal pressure range and lower. This is due to: 1) the horizontal orientation of the flow phantom where higher elevation branches experienced less head backpressure, and 2) the long length and varied si zes of the outflow pipes. The absolute and relative flow rates of the major portal vein branches (LPV, RAPV, RPPV) are plotted as a function of portal inlet pr essure in Fig. II.22.

PAGE 109

88 Figure II.22 Major Portal Branch Flow Rates The Right branches of the portal vein carry nearly 90% of the blood flow with 70% going to the anterior of the liver and 17% to the posterior region. Again, the Left branch distributes just above 10% to a smaller parenchymal mass. CFD Simulation Results and Discussion A streamline plot of the CFD simulation is shown in Fig. II.23.

PAGE 110

89 Figure II.23 Velocity Profile Simulation of Phantom Pattern Mesh The high velocity red flows occur where the mesh was manually edited for smoothing or the inlet outlet pipe joints creating un-natural surface and cross-sectional area transitions. Comparison of Thin Walled Phantom and CFD Simulation Results and Discussion The flows of the major portal vein branches of the wet phantom and the results of the CFD simulation are co mpared in Fig. II.24.

PAGE 111

90 Figure II.24 Comparison of CFD Es timates with Phantom Measurements The portal phantom flow measurements and CFD simulation tracked well for branch #3. The branch #5 comparison track well excepti ng the shift in flow from the LPV to the RAPV at low pressure/flow data shown in Figs. II.22-23. The branch #4 comparison tracks well with shows a 15% difference that may be explained by the measurements being taken in a horizontal phantom orientati on, which places the branches at elevations +/1” (2.5 cm) of the portal elevations which places the upper outlets at a slightly lower back pressure than the deeper outlets. This physical orientation is not accommodated in the setup of the simulation.

PAGE 112

91 Aim 2 Obtain an ROI from a 3D Image Volume from Clinician Input ADP ROI Results and Discussion The ADP ROI capture algorithm worked re liably but the process and results are not shown since ADP was not used for al gorithm verification. The algorithm was migrated into the implementation of VAS with minor modification. VAS (C/C++/OpenGL/Qt) ROI Results and Discussion [Ultrasound Data] The VAS ROI capture method was succe ssful in capturing random ROIs over hundreds of uses during the algorithm verificati on phase. A successful capture process is shown in Figs. II.25-27. The Lateral view RO I selection box set by th e user is shown in Fig. II.25. Figure II.25 ROI Selection Box (Lateral View)

PAGE 113

92 The Elevation view ROI selection box set by the user is shown in Fig. II.26. Figure II.26 ROI Selection Box (Elevation View) The completed ROI selection is shown in Fig. II.27 with a captured and highlighted vessel.

PAGE 114

93 Figure II.27 ROI Selection Completed The ROI selection process requires the clinicia n to size and position the selection box in both views, a task that was accomplis hed in under five seconds per view. Aim 3 Autonomously Segment Va sculature and Compute Centerline Segmentation Development ADP Results and Discussion A typical capture of a vessel crosssection is shown in (Fig. II.28).

PAGE 115

94 Figure II.28 Edge, Center, and Elliptical Approximations of Vessel Cross-section The Contour (CNT) function properly identified the perimeter of the vessel cross-section which was plotted as a blue overlay; the Center (CEN) function properly compute the cross-section center plotted as a red circle overlay; and an el liptical representation of the vessel cross-section is plotted as a red ellipse overlay. The 9 x 9 x 9 voxel Gaussian smoothing of the 3D Doppler data (Red-left, BlueRight) performed by the ADP Surface (SUR) fu nction is shown in Fig. II.29. Figure II.29 Gaussian Smoothing of Doppler Data

PAGE 116

95 The resulting surface was unacceptably rough an d the processing time took 10 minutes to complete, which was not in lin e with the required TTR. VAS (C/OpenGL) Results and Discussion [Visual Human Data] A successful vessel capture is provided in Fig. II.30 where the edges of the vessel are identified by red dots, the horizontal centerlines ar e represented in green dots, the mass centroid is shown as a white dot and the maximum vessel dimensions are shown by a grey bounding box. Figure II.30 Outline and Center Computed for Vessel Cross-section A successful segmentation of two close proxim ity vessels with horiz ontal and vertical overlap is shown in Fig. II.31.

PAGE 117

96 Figure II.31 Typical Vessel Overlap and Proximity The full point clouds for the portal vein, he patic vein and liver capsule are shown in red, green and blue colors in Fig. II.32. Figure II.32 Segmented Point Cloud of He patic Capsule and Portal/Hepatic Veins

PAGE 118

97 The detail obtained by this segmentati on method provided much more detail at finer branching levels than obt ained from the matrix method used in creating the mesh for the portal vein phantom pattern [58]. The mass centroids of vessel cross-sections are plotted as white dots in Fig. II.33. Figure II.33 Centerline Approximation While the centers of vessels sliced across the diameter provided a reasonable centerline track for a snaking approach, the technique does not yield useful information when vessels are sliced along their length.

PAGE 119

98 VAS (C/C++/OpenGL/Qt) Results and Discussion [Ultrasound Data] The segmentation function developed for th e clinical application successfully identified up to four separate vessel obj ects within the ROI and color coded them as shown in Fig. II.34. Figure II.34 Multi-Vessel ROI Capture The success of segmentations was de pendent on the surface smoothness of the Doppler point clouds since surface noise on some vessel A (in close proximity to a vessel B) might fall within the voxel capture range of vessel B, creating an unwanted connection and a failed segmentation. As an illustration of how surface noise reduces the effective spacing and caused failed segmentations, two noisy and widely spaced phantom tubes are shown in Fig. II.35, where the segmentation would have failed had the capture range been larger or the gap betw een the tubes been smaller.

PAGE 120

99 Figure II.35 Surface Noise in Color Doppler Point Clouds Surface noise comes from Segmentation Error Source #1, which is insufficient echogenicity and homogeneity in the test fl uid causing high spatial noise in the Color Doppler data. Using the Color Doppler acqui sition technique described in the methods section with water/cornstarch and water/gl ycerol, the resulting Doppler surfaces were unusable “chunky/lumpy” chains of “color bl obs”. Useable Doppler surfaces were obtained with higher gain at the expense of th e Doppler color data ex ceeding the internal diameter of the lumen, an effect known as “blooming”. The water/corn starch test fluid provided high echogenicity for the short period of time after introduction into the flow but settled out of so lution quickly, making stable long term measurement series difficult. The corn starch particle size and homogeneity varied greatly and caused many “lumpy” artifacts in the rend ered color Doppler surfaces caused by the large more reflect ive scatterers which flowed through the beam faster than the Automatic Gain Control feature of the ultrasound cart could respond. The Segmentation Ga p

PAGE 121

100 water/glycerin mixture concentrations used were more stable, but, exhibited a much lower echogenicity (requiring higher Doppl er gain) and regions of non-homogenous mixtures which resulted in flashes of th e Doppler color outsid e the lumen, with the appearance of a lensing effect As a more echogenic and stable blood analogue was not available, it was decided that using Colo r Doppler surfaces which were obviously “blooming” outside the vessel lumen would be sufficient for testing the functional intent of the Aims and VAS capabilities with the understanding that the diameters and subsequently calculated areas and flows woul d be substantially over-estimated. Final verification will require use of blood, micro-bubbles or nano-par ticles that can provide an adequate, homogeneous and consistent echoge nicity and sub-luminal Doppler data. VAS Vessel Centerline Results and Discussion The centerlines of the segmented and highlighted vessel segments in Fig. II.34 were successfully computed and displayed as wh ite dots at the centers of the partitioned compartments and connected by white lines. An additional successful centerline of a long 2” (5 cm) curved vessel is shown in Fig. II.36 in top and side views.

PAGE 122

101 Figure II.36 Long Curved Vessel Test (S0324-65R) Centerline Error Source #1 is rough D oppler surfaces caused by insufficient echogenicity and homogeneity in the test fluid. Since the local voxel “mass” used to find the center of each compartment includes the voxe ls of the Doppler surf ace, this technique is sensitive to rough Doppler surfaces. While the centerlines shown in Figs. II.34 and 36 are well defined and curvilinear, the centerline segment shown in Fig. II.37 has been clearly influenced by the “lumpy” surf ace in the neighborhood of the calculation, specifically, the centerline segment was placed within the volume of spatial noise.

PAGE 123

102 Figure II.37 Centerline Direction Error Due to Surface Noise The error source can be explained as a two dimensional (2D) single ended case with a single line of voxels overlaying a vessel di ameter. In Example 1 of Table II.1, the addition or subtraction of the same number of noise voxels at both ends of the line does not displace the center from its nominal posi tion since symmetry is maintained. In Example 2, the addition of an equal number of noise voxels to one end, as are subtracted from the other end, displaces the centerline from its nominal position by that number of voxels in the direction of the added no ise voxels. In Example 3, the centerline displacement error due to an imbalance of voxe ls added/subtracted at either end is one half the number of voxel difference in the direction of th e imbalance. Incorrect Centerline Correct Centerline

PAGE 124

103 Table II.1 Centerline Error from Surface Noise Example End 0 Displacement End 1 Displacement Center Displacement 1a + n n 0 1b n + n 0 2a + n + n + n 2b n n n 3a 0 +/n +/n/2 3b -/+ n 0 -/+ n/2 In reality, both of the voxels representing the ends of the centerline segment can be displaced by surface noise. In the 3D case, the positional error of the center due to noise is the combined displacements in the x, y, z dimensions in th e direction of the net imbalances as related by Eq. II.7. x 100 Cnom = Nominal Center Position Cdisp = Displaced Center Position Equation II.7 Centerline Error from Noise

PAGE 125

104 Aim 4 Autonomously Computed Segmented Vessel Diameters and Cross-sectional Areas VAS Diameter Results and Discussion Color Doppler The computed centerline segments and the cross-sections defined by the Dot Product method (Eq. II.3) for two random ROI captures of the 3/8” (0.952 cm) tubing in the net flow phantom are shown in Figs. II.38-39. Figure II.38 Color Doppler Cross-section Capture (S0324-65B-3) Figure II.39 Color Doppler Cross-section Capture (S0313-70B-5)

PAGE 126

105 The diameter errors from the cross-secti ons at the top (Cente rline Segment 0) and bottom (Centerline Segment 1) of five random ly positioned ROIs within three separate image volumes are shown in Fig. II.40. Figure II.40 Color Doppler Diameter Measurements

PAGE 127

106 The statistics for the data presented in Figure II.40 are prov ided in Table II.2. Table II.2 Diameter Error Statistics Image Error Mean Error S0324-65B 93% 8% S0324-65R 124% 24% S0313-70B 50% 9% Since the ROIs were captured from known overs ized Color Doppler da ta as discussed in the methods section, the mean of the diameter errors are the expect ed overestimates of the actual tube size. Diameter Error Source #1 is the use of known oversized (blo omed) Doppler data as discussed above. The errors in the means of the ROI samples of the images is due to the operator adjustment of the gain to get a smooth Color Doppler Data volume, an issue expected to be resolved with use of more echogenic test fl uids. The mean error of the S0313 samples were better than the S0324 captures by at least 50% due to a more echogenic corn starch contrast concentration at the time of te st which allowed the use of a lower gain and resulted in Color Doppler data closer to the actual lumen diameter. The low deviation in the S0324-65B samples suggest that the diameter of the Color Doppler data volume tube was near circular and cons istent along the sampled length. The higher deviation in the S0324-65R samples (visible in the End 0/1 differences of samples 2-4) was due to the ROI cutting the Color Doppler Data where there was visible diameter taper due to curvature in the Color Doppler data volume at both ends as shown in Fig II.36.

PAGE 128

107 Diameter Error Source #2 is error in the computed centerline direction as discussed above. If the center line is not collinear with th e vessel then the Dot Product process of finding vessel cross-section will ma ke an elliptical cut through the nominally circular vessel. The centerline error angle is computed between the nominal direction of the centerline vector A and the incorrect di rection vector B as shown in Eq. II.8. A = Nominal centerline direction B = Displaced centerline direction = Angle between nominal and displaced centerlines Equation II.8 Centerline Error The effects of centerline error on computed vessel diameter are shown in Table II.3 where errors above 30 would be unacceptable. Table II.3 Diameter Errors from Centerline Centerline Error () Radius/Diameter Error (%) 0 0.0% 10 1.5% 20 6.4% 30 15.5% 40 30.5% 50 55.6%

PAGE 129

108 For a single dimensional example with a typical image resolution of 0.5 mm/voxel, a typical centerline le ngth of 10 voxels (5 mm), a typical centerline noise displacement of 2 voxels (1 mm) from surface noise, and, a t ypical vessel diameter of 10 mm (20 voxels) from Table I.1: the centerline error angle would be 11.3, and, the radius/diameter error would be 1.5% (0.15 voxels). Examination of Table II.3 indicates that it would take a 6 voxel deviation for each 10 voxels of centerlin e run (30) to achieve a 15.5% error in radius/diameter (visibly illustrated in Fig. II.37 as a nearly 45 error), suggesting that the diameter determination is relatively insensitive to centerline error. Diameter Error Source #3 is natural vesse l cross-sectional eccentricity shown in Fig. II.41, which suggests that th e initial method of using the average of the three longest radii, while fast and simple, may not be su fficient for handling out-o f-round vessels since the longest three radii will be on the long axis and cause overestimation of the area. The impacts of this error are discussed be low in the Area Computation section. Figure II.41 Color Doppler Diamet er Error Eccentr icity (S0313-70)

PAGE 130

109 A new diameter estimation method is under consideration based on the prototype developed in the ADP (Fig. II.28) and is e xpected to provide improved accuracy which will approximate the cross-section as an ellipsoid from which major and minor axes can be calculated. The elliptical algorithm is expect ed to be slower since the ellipse must be a best “fit” of the vessel shape which is more computationally expensive. Diameter Error Source #4 comes from irre gular voxel capture in the cross-section as shown in Figs. II.41 and as highlighted in the side view provided in Fig. II.42 (upper), where the gap in the pattern ma y occur at the vessel edge where radius measurements are taken.

PAGE 131

110 Figure II.42 Color Doppler Diameter Error Pattern (S0320-34)

PAGE 132

111 This patterned effect comes from the f act that the test for voxels on a surface normal to the centerline defined by the Dot Pr oduct method in Eq. II.3 is done in floating point math to accommodate the varied sub-millimeter dimensions of the voxels (cuboids, not cubic). Since it is rare in floating point math to ha ve a perfectly exact comparison with zero (0.000) to 16 digits of mantissa and an exponent range of +/383, the test allows for voxels that are within +/.00001 of zero. Dependi ng on the size of the vessel and angle of the centerline to the Cartesian directions, some voxel centers may not fall within the capture tole rance, or, alternatively, many voxels may be captured,. Centerline directions closer to ninety (90) or zero (0) degrees will capture more voxels (Fig. II.42 lower) since cuboids have a high center pitch de nsity in the x, y, z di rections, and a lower pitch density near 45 degrees from any Cartes ian defined plane. Small diameter vessels exhibit few voxels highlighted in the captured cross-sections typically with the patterned effect illustrated above as there are few voxels available to meet the tolerance described above, where cross-sectional captures of la rger diameter vessels may highlight many voxels due to the abundance of voxels meeting the tolerance. In cases of very thick cross sectional captures, the longest radius measurem ents are to the edges of the disk which are not in the perfect cross-sectio nal plane. This error could be reduced by adjusting the capture tolerance with the angle to the Ca rtesian directions, sp ecifically a tighter tolerance near zero and ninety degrees a nd a wider tolerance near 45 degrees. The pattern effect would not influence the curre nt diameter finding al gorithm as it will find the most distant voxels in the entire circum ference to work with, ignoring pattern gaps.

PAGE 133

112 Diameter Error Source #5 is the relati onship between the noise floor of the ultrasound cart and the intensity threshold us ed to determine the position of physical surfaces. The algorithm developed to detect a nd noise filter the Doppler surface uses an intensity threshold of one (1) (a ll voxel intensity values above 0) to qualify the start of a surface to ensure that the largest inner diameter of the vessel is detected. The risk in this approach is that if there is spatial noise in the ultrasound data, this low threshold will create a surface envelope based on noise. To test this relationship, a higher threshold value was used with an anatomical image (Fig. II.43 left) to see if there was an improvement in surface definition or whether an erosion of features would occur, the latter being the result as s hown in Fig. II.43 (right). Figure II.43 Surface Detection Threshold Set to 1/128 and 25/128

PAGE 134

113 Since use of a higher threshold value may erode legitimate surfaces, a future method might include the parallel detection of the surface within the 2D BW volume understanding that BW vessel de tection can be more difficu lt due to lower contrast. Diameter Error Source #6 is an incomple te capture of the cr oss-section at low angles as seen at the upper and lower margins of the ROI in Fig. II.38-39. The current method of computing diameter is insensitive to this error for circul ar vessels since it needs only half a cross-section to determin e a radius. The truncation at the boundary occurs because the clinician sets the ROI to avoid other anatomy and noise in the image and the voxels outside the ROI are not include d in the segmentation. Picking a centerline segment more central to the ROI interior w ould be a compromise at the expense of the velocity not being calculated at the ROI control volume bounding surface which may be important to the clinician or in avoidance of other anatomy. This error source would be a large issue for area calculation methods which sum all voxels. Diameter Error Source #7 can be seen in Figs. II.36,44 where the ends of the curved section of the net phantom tube appear to taper from the nominal tube diameter to zero diameter as an artifact of the Color D oppler sweep and sensitivity to flow. ROIs which include the tapering vessel will give erro neously low diameter values as shown in the S0324-65B results in Table II.2.

PAGE 135

114 Figure II.44 End Tapering of Color Doppler Surfaces (S0324-65BR) Power Doppler The inlet (left) and outlet (r ight) diameter estimates from sample S0407-41 in Fig. II.45 are 1.04 cm and 1.24 cm respectively, 11% and 32% over the reference diameter of 3/8” (0.952 cm). Figure II.45 Power Doppler Diamet er Measurement Tubing (S0407-41)

PAGE 136

115 Where the Color Doppler acquisitions requi red excessive gain to acquire a smooth surface (although outside the tubing lumen), Power Doppler provided a smoother surface within the lumen with the water based fluids using the 2D and Doppler adjustment procedure discussed in the methods. The expl anation of this measurement difference is that Power Doppler utilizes th e amplitude of the returning si gnals to infer the relative magnitude of scatterer velocities in the beam path that gives it more sensitivity since amplitude losses can be compensated with amp lification. In contrast Color Doppler uses the spectrum of frequencies returned to infe r relative velocities, where frequency is the time rate of change of phase. Since the pha se shifts many be slight, and improving phase resolution is not simple, Color Doppler is ge nerally not as sensitive as Power Doppler. Unfortunately, Power Doppler cannot be used as the sole acquisition mode since it does not provide an absolute measur ement of velocity. A hybrid approach to acquisition could be implemented on the ultrasound cart using a Power Doppler sweep to obtain diameters and a Color Doppler sweep for acquiring velocity The primary issue with this two mode approach is that any physical movement in the transducer during or between the sweeps would cause significant physical mis-registrati on in the Power and Color Doppler image volumes. If implemented externally to th e cart, it would also be a time consuming procedure for the clinician, would depend on cap turing two separate files, and would not meet the TTR.

PAGE 137

116 Area Results and Discussion Manual calculations were performed to ve rify that VAS was correctly computing inaccurate areas from the inaccurate diameter s reported above. The vessel eccentricity observed in the cross-sectional capture results discussed in sections above suggest that the method of calculating area described by E q. II.4, although TTR compliant, will not be adequately robust for clinical use. The el liptical approximation me thod discussed in the diameter results for improving vessel diamet er estimation with out-of-round vessels would also provide area as an output. The area for elliptical vessels and off-centerline elliptical cuts of ci rcular vessels is com puted from Eqn. II.9. Equation II.9 Elliptical Area The area error relative to a perfect circle is presented in Table II.4 as the major and minor axes are varied as a ratio of each other.

PAGE 138

117 Table II.4 Area Errors from Diameter Error Axes Ratio (A/B) Area cm2 Error (%) 1 3.14 0% 1.1 3.46 10% 1.2 3.77 20% 1.3 4.08 30% 1.4 4.40 40% 1.5 4.71 50% Interestingly, the elliptical error differences over circular are lin ear and proportional to the ratio of the axes. Interpolating Tables II.3 and II.4, it would take a 25 centerline error to create a 10% error in both diameter a nd area, thus area erro r is a stronger function of diameter error than of centerline error. Aim 5 Autonomously Compute Angle Corrected Velocities for Segmented Vessels VAS Angle Correction Results and Discussion The transducer origin location was succe ssfully located by visual inspection in Fig. II.46 (left) despite the high spatial edge noise shown in Fig. II.46 (right).

PAGE 139

118 Figure II.46 BW Image Envelope and Computed Transducer Origins The angle between the transducer beam lines (grey) and the segments of the centerline (white) of a segmented vessel (red ) are shown in Fig. II.47, zoomed out to include the transducer origin in the left imag e and zoomed in to a close view of the vessel segment on the right.

PAGE 140

119 Figure II.47 Transducer Beam Lines Velocity Test Calibration Re sults and Discussion A comparison of volume flow rate s derived from random Mass/Time measurements and manual PW Doppler Angle Co rrected velocities is shown in Fig. II.48 with +/5% error bars.

PAGE 141

120 Figure II.48 Comparison of Mass/Time and Manual Angle Correction The results for the random samples are within 5% of each other for three of the four samples. The Reynolds numbers from each of the samples were 2003.7, 1204.2 4263.9, and 4772.7 respectively, predominantly a function of the velocities as the diameters fell within a 20% range. The first two samples are below the <2300 suggested threshold for laminar flow. The second two samples are between the laminar threshold and the suggested threshold of >6000 fo r full turbulent flow, suggesti ng that the flow may be a combination of laminar and turbulent [41]. The gap in S29 results may come from the following two error sources. Flow Error Sour ce #1 is the Mass/Time method involve the weighing (+/10 g) and timing (+/1 s) fr om the balance and stopwatch used for the measurements. Flow Error Source #2 is the graphically and visually inaccurate procedure for manual angle correction: 1) th e visual placement of the center of the Sample Volume (SV) in the center of th e 2D representation of the vessel on the

PAGE 142

121 ultrasound cart screen by the clin ician; 2) the cl inician selected size of the SV relative to the visible lumen of the vessel on screen; 3) the clinician visual setting of the cursor in the direction of the vessel which may be straight or curving. Raw Velocity (Cross-sec tional Mean and Centerline) The uncorrected Color Doppler velociti es from three image volumes and five randomly selected ROIs are shown in Fig. II.49.

PAGE 143

122 Figure II.49 Uncorrected Velocities

PAGE 144

123 With the transducer set at 45 to the phantom tube as shown in Fig. II.50, the samples of 65B-1, 65R-5, and 70B show the expected step-wise “staircase” increase in beam relative velocity from the shortest beam path (low est relative Doppler ve locity at centerline Segment 0 End 0) to the longest beam path (highest relative Doppler velocity at centerline Segment 1 End 1). Figure II.50 Location of Velocity Samples The randomly selected ROIs capturing thes e samples are of random size, thus the samples with smaller velocity variations between segment 0 and segment 1 (70B-1, 65B5) were captured over shorter lengths of the phantom tube, where ROIs capturing more tube exhibited much larger spreads of velocity (65R-4). The statistics for the samples shown in Fig. II.49 are provided in Table II.5.

PAGE 145

124 Table II.5 Uncorrected Velocity Statistics Image/Group Mean (cm/s) S0324-65B Overall 22.35 1.42 Seg 0 21.32 1.00 Seg 1 23.38 0.94 S0324-65R Overall 14.64 2.19 Seg 0 12.85 1.21 Seg 1 16.43 1.31 S0313-70B Overall 12.11 1.89 Seg 0 10.79 1.05 Seg 1 13.42 1.60 The step-wise velocity profile is confirmed by a step wise relationship among the means: Segment 0 means for each image in Table II.5 are the lowest, the Segment 1 means are the highest, and the overall m eans are between the two. Some of the sampled velocities do not e xhibit perfect stair step relationships where the End 0 velocity is lower than End 1 (65B-2,4,5) within a segment, the reverse of what was expected. This is can be explaine d by circular turbulent flow in the Doppler beam path where fluid on one side of a segm ent sized vortex could have a velocity below the mean, and, the velocity on the other side could be above the mean, resulting in erroneous end velocities due to the counter flow. To gain in sights into the flow profile for this image, the Reynolds numbers for each image are shown in Table II.6 as computed using Eq. II.10 using the density of water at 1000 kg/m3 and dynamic viscosity of 1E-3 kg/ms from Table I.2, the Color Doppl er data average diameter from the sample data, and the manually angle correc ted velocities for each sample.

PAGE 146

125 Equation II.10 Test for Laminar Vs Turbulent Flow Table II.6 Reynolds Numbers for Tested Images Image Doppler Diameter (D) (m) Velocity ( ) TAMV (m/s) Reynolds Number Nearest Limit Flow Type S0324-65B 0.0185 0.32 5920 ~6000 Turbulent S0324-65R 0.0214 0.32 6848 ~6000 Turbulent S0313-70B 0.0143 0.23 3289 ~2300 Laminar The 65B and 65R images showed the most samp les with End 0/1 velo city revers als (only one step-wise sample each) and were taken under flow conditions determined to be turbulent in Table II.6. The 70B image was sampled under laminar conditions and showed only one sample with a slightly reve rsed End 0/1 velocity. Since the Reynolds number computed for a nominal portal vein was calculated to be a very laminar 397.5 (<2300) in Eq. I.5, future in-vivo results are ex pected to look similar to the 70B images samples rather than the 65B/R samples whic h were taken under flow conditions that would not be found in normal physiology. Velocity Error Source # 1 is the small range of velocity/color variation over the length of the capturable vessels. The deviati ons of the uncorrected velocities in Table II.5 are very low and call attent ion to the roughly 5cm/s range of velocities over the full length of the longest capturable Doppler data volumes using the widest possible field of view and imaging depth. This small numeric velocity change across the captured vessels is related to a small color change in the vi sible true color images on the ultrasound cart.

PAGE 147

126 In Fig. II.51, an image of a phantom tube is captured at 90 to the transducer with a phantom flow set just below max scale to show the full range of color /velocity transition (right to left) of fast flow towards the transduc er (white) to red to zero (black) to blue to fast flow away from transducer (white). Figure II.51 Manual Angle Correction and Color Scale (Cart) The same image capture is shown in Fig. II.52 as rendered by VAS to show accurate color/velocity reproduction. Figure II.52 Manual Angle Correction and Color Scale (VAS)

PAGE 148

127 The hue (velocity) of the red Color D oppler in Fig. II.53 can be seen varying slightly along the length of th e tube in from darker to li ghter to darker as the tube direction changes the direction of the flow relative to the beam path (origin off the right edge of the image). Figure II.53 Doppler Velocity True Color Variation (S0324-65) Angle Corrected Velocities (Crosssectional Mean and Centerline) S0324-65B Results Random ROIs were selected in image S0324-65B (Fig. II.39) and the crosssectional mean and centerline angle corrected velocity errors for both ends of the captured vessel are shown in Fig. II.54. Th e Maximum Color Doppler velocity scale set for these readings was 38.5 cm/s. The angl e corrected velocities obtained manually from

PAGE 149

128 several locations along the net flow phantoms tube from the cart using Pulse Wave (PW) Doppler were 32 cm/s (TAMV) and 74 cm/s (TAPV). The cross-sectional mean velocities were compared to the TAMV and the centerline velocities were compared to the TAPV to obtain error magnitudes. Figure II.54 Straight Tube Velocity Errors (S0324-65B)

PAGE 150

129 The velocity errors computed for th e samples are shown in Table II.7. Table II.7 Straight Tube Velocity Errors (S0324-65B) Error Mean Error Vel AC Centerline -35% 16% Vel AC Cross-secti onal Mean -46% 10% The angle corrected centerline velocity estimates averaged 35% below the measured TAPV from the ultrasound cart with a modest deviation. The angle corrected cross-sectional mean velocity estimates av eraged 46% low, sugge sting that a better method of computing the cross-sectional mean velocity is needed. With a Reynolds number of 5920 indicating tu rbulent flow for this imag e, the centerline and crosssectional mean estimates were expected to be similar since the centerline velocity is from a voxel that is included in both calculations. Velocity Error Source #2 is an artificially low cross-sectional mean velocity caused by incomplete capture of the vessel cros s-section as discussed in Diameter Error Source # 4 where not all of th e cross-sectional v oxels are captured due to the capture tolerance and in small vessels. This prevents uncaptured voxels from contributing to the mean. Velocity Error Source # 3: The centerlin e velocities retrieved from the Color Doppler Data may not be perfectly on the centerline as discusse d in Centerline Error Source #1.

PAGE 151

130 Velocity Error Source #4 may be di fferences in the manual 2D and the autonomous 3D angle correction methods. In the manual method, the clinician orients the transducer to slice the vessel of interest in a 2D plane, then lo cates the sample volume and angle corrects as discussed above. This procedure happens in a single plane defined by the depth and lateral transducer view di mensions (x y plane in VAS). Figure II.55 shows the VAS 3D angle correction geometry from the transducerÂ’s perspective. Figure II.55 Straight Tube Blue Top View (S0324-65B) Both of the centerline se gments are oriented in the same (x) direction, but are offset from the x y plane. In 2D practice, the clinician w ould tip the transducer to eliminate the z axis distance (with poor accuracy). In the 3D case, cos is calculated from Eq. II.6, which accommodates the fact that the vertices of the centerline segment and transducer beam have x, y, and z coordinates. In a 3D im age volume of hepatic parenchyma, there are many vessels of varied sizes which are orient ed in random directions (Fig. I.7) which

PAGE 152

131 confounds the concept of orienting or identifyi ng a 2D plane in which to do a 2D angle correction that would acc ommodate all vessels. In the 2D method, the Doppler velocity is adjusted by one Cosine, in the 3D method there is an implicit Cosine correction for the angle the 2D plane is tilted out of the x y plane and additional Cosine correction for the angle the centerline makes with that 2D plane. The second cosine correction may explain why the VAS velocity may not be imitating th e manually acquired value for the cart. As an example, a centerline segment which is 45 out of the x y plane and 45 rotated from that plane would be corrected by 0.707 x 0.707, a much larger correction than a 2D correction of 0.707.

PAGE 153

132 S0324-65R Results To illustrate the effect of disparate centerline directions on the estimated angle corrected velocity, a series of random ROIs were set along a hi ghly curved tube in the net flow phantom in S0324-65R as shown in Fi g. II.56. The maximum color Doppler scale was set to 38.5 cm/s and the angle corrected velocities obtained manually from several locations along the net flow phantoms tube fr om the cart using Pulse Wave (PW) Doppler were 32 cm/s (TAMV) and 74 cm/s (TAPV). Figure II.56 Curved Tube Top View (S0324-65R) The velocity errors taken from the ROIs are shown in Fig. II.57.

PAGE 154

133 Figure II.57 Curved Tube Velocity Errors (S0324-65R) The curved tube velocity error st atistics are shown in Table II.8. Table II.8 Curved Tube Velocity Errors (S0324-65R) Error Mean Error Vel AC Centerline -54% 11% Vel AC Cross-sectional Mean -43% 18%

PAGE 155

134 The errors for the curved tube image were similar to those for the straight tube in centerline velocity estimates. There were higher End 0/1 error differences in several samples (1-3) due to the varied centerline directions along the tube curvature. The Reynolds number of 6848 suggest s turbulent flow, was expected to be higher in the curving tube, and may explain the slightly higher cross-sectional mean error and deviation over the straight image. Doppler Accuracy Velocity Error Source #5 is the accurate angle of insonation range for Doppler, commonly cited as 30 to 60. Angles below 30 risk inaccuracy due to the sound waves being reflected by the vessel surface and low sens itivity in the Cosine as it nears a value of 1.0 at 0. Angles above 60 involve the Co sine insensitivity as it reaches a value of 0.0 at 90, and, less relative frequency/velocity shift for the Doppler to accurately detect. Because of these issues, there is a finite length of each vessel that falls within the accurate range of angles in an image of ma ny randomly oriented vessels. The “zones of accuracy” for each vessel could be identified in future work through non-TTR compliant processing of the captured centerlines agai nst all beam angles in the 3D sweep. Impact of Rough Color Doppler Surface on Velocity Results and Discussion The Series phantom was used for data capture and a True Color image of the three tubes is shown in Figs. II.58.

PAGE 156

135 Figure II.58 Series Phantom All Tubes The flow was set to use the full velocity scale of the 1/4” tube segment (middle tube) (slight aliasing visible), forcing the colo r velocity in the upstream 3/8” tube (right) the dark (low) end of the scale, and forci ng the color velocity in the downstream 1/8” tube over the aliasing limit for the selected sc ale. (When the flow velocity was set below aliasing for the 1/8” tube, the velocity in th e 3/8” tube was so lo w (black) that Color Doppler data could not be captured within the tube.) A Random ROI sample of the image in Fi g II.58 is shown in Fig II.59. Imaging the 3/8” tube alone, the maximum Color Doppl er scale velocity was 33.9 cm/s and angle corrected velocities obtained manually from severa l locations along the net flow phantoms tube from the cart using Pulse Wave (PW) Doppler were 23 cm/s (TAMV) and 39 cm/s (TAPV).

PAGE 157

136 Figure II.59 Rough Surface Capture Top View (S0313-70B-5) The velocity errors computed from five ra ndom ROI captures are shown in Fig. II.60. Figure II.60 Rough Surface Velocity Errors (S0313-70B)

PAGE 158

137 The statistics of the velocity errors are shown in Fig. II .60 are shown in Table II.9. Table II.9 Rough Surface Velocity Errors (S0313-70B) Error Mean Error Vel AC Centerline 14% 55% Vel AC Cross-sectional Mean 21% 65% The diameter and velocity errors for this image were the best of the three images for both the centerline and the cross-section mean estimates on seven of ten samples with the wide deviations caused by outlier end valu es driving the high devi ation. If samples 3 and 4 are omitted from the error statistics, th e error and deviations become a much more reasonable -15% and 35% for the cross-se ctional mean, and, -17% and 19% for the centerline respectively. Th e Reynolds number was 3289 for this image and was the closest to the threshold for laminar flow sugge sting that the algorithm s will be capable of accurate flow estimation under normal laminar hepatic flow conditions. Aim 6 Autonomously Compute Volumetric Flow for Segmented Vessels Flow Results and Discussion An ROI capturing a pair of net flow phantom tubes is shown in Fig. II.61.

PAGE 159

138 Figure II.61 Net Flow Phantom, Straight-Curved Tubing (S0324-65) Flow is computed from the angle correct ed velocity, and, the computed area, per Eq. I.1. The inlet and outlet velocities, ar eas and flows computed by VAS are provided in the ROI data display area (lower right) along with the total volumes for the Red and Blue Doppler in the ROI and also w ithin the entire image volume. Given the use of Color Doppler images with known oversized diameters and having exposed a number of diameter and veloci ty error sources, reporting of flow results compared to the measured flows would not be appropriate. However, since the net flow phantom was devised to facilitate the quan tification of net flow through the ROI as a

PAGE 160

139 control volume, net flow was tested and the re sults appear in the lower right corner of Fig. II.61. The flows through the ROI balance within 1.7 cc/s, an error of 2.5%. This result suggests that the diameters and veloci ties, while known to be wrong, were equal in the opposite flow directions with offsetting erro rs. Five samples of net volume flow for individual straight and curved tubes are shown in Fig. II.62.

PAGE 161

140 Figure II.62 Net Flow Errors

PAGE 162

141 The curved tube net flow statistics are shown in Table II.10. Table II.10 Net Flow Velocity Errors Error Mean Error S0324-65B 7% 6% S0324-65R 2% 8% S0324-70B 15% 6% The net flows computed to 15% of the aver age inlet/outlet flows over all images. The means and deviation of the net flow erro rs were admirably small despite the known oversized Color Doppler data volume s and previously discussed errors. Time to Response On the Laptop PC used for development th e clinical application met the TTR for cloud based images which rotate in real ti me without delays from mouse and keyboard control inputs. The ROI capture process was real time as gated by clinician input. The ROI voxel identification, cente rline computations, and angl e corrections take under 20 seconds for ROIs containing up to four objec ts which would be typical of near tumor captures. The time required to import and cloud detect the initial imag e file is under 30 seconds, a time dominated by the use of intermedia te binary array files as an interface to the MATLAB file extractor. This file impor t time will be reduced in a future version with direct access to DICOM files.

PAGE 163

142 CHAPTER III FULL TEXT OF AIM 1 PHANT OM MATERIALS (PUBLISHED) Title An investigation of industrial molding com pounds for use in 3D Ultrasound, MRI and CT imaging phantoms. Authors and Affiliations Bryan E. Yunker1, Dietmar Cordes PhD2, Ann Scherzinger PhD2, Gerald Dodd MD2, Robin Shandas PhD1, Yusheng Feng3 PhD, Kendall S. Hunter PhD1 1 Department of Bioengineering, University of Colorado, Denver, Colorado Department of Bioengineering University of Colorado at Denver 12700 E. 19th Ave, MS 8607 Aurora, CO 80045-2560 Phone: 303-724-5893 Fax: 303-724-5800 E-mail: bioengineering@ucdenver.edu 2 Department of Radiology, University of Colorado Hospital, Denver, Colorado Department of Radiology University of Colorado SOM Mail Stop L954 12401 E. 17th Ave. PO Box 6510 Aurora, CO 80045 Phone: (720) 848-6608 FAX: 720-848-7315 E-mail: paula.larrabee@ucdenver.edu 3 College of Engineering, Departme nt of Mechanical Engineering University of Texas San Antonio Department of Mechanical Engineering Mail Stop: AET 2.332 One UTSA Circle San Antonio, TX 78249-0670 Phone: 210-458-6479 FAX: 210-458-6504 Email: yusheng.feng@utsa.edu

PAGE 164

143 Abstract Purpose: This study investigated the Ultrasound, MRI and CT imaging characteristics of several industrial casting and molding compounds as a precursor to the future development of durable and anatom ically correct flow phantoms. Methods: A set of usability and performance crite ria was established for a proposed phantom design capable of supporting liquid flow during imaging. A l iterature search was conducted to identify the materials and methods previously used in phantom fabrication. A database of human tissue and casting material properties was compiled to facilitate the selection of appropriate materials for testing. Several industrial casting materials were selected, procured, and used to fabricate test sample s that were imaged with Ultrasound, MRI and CT. Results: Five silicones and one polyurethane we re selected for testing. Samples of all materials were successfully fabricate d. All imaging modalities were able to discriminate between the materials tested. Ultrasound testing showed that three of the silicones could be imaged to a depth of at least 2.5 cm (1 in). The RP-6400 polyurethane exhibited excellent contrast and edge deta il for MRI phantoms and appears to be an excellent water reference for CT applications The 10T and 27T sili cones appear to be usable water references for MRI imaging. Conclusions: Based on study data and the stated selection criteria, the P-4 silicone provided sufficien t material contrast to water and edge detail for use across all imaging modaliti es with the benefits of availability, low cost, dimensional stability, non-toxic, non-fl ammable, durable, cleanable, and optical clarity. The physical and imaging differences of the materials documented in this study may be useful for other applications.

PAGE 165

144 Keywords: Imaging, Phantom, Ultrasound, MRI, CT Purpose This study investigated i ndustrial casting and molding compounds for use in a proposed hepatic flow phantom design for Ultrasound, MRI and CT imaging. Many of the phantoms cited in literature are for “sta tic” applications [59-63] that do not involve flowing fluids even if the phant om contains a fluid as a contra st or a calibration reference. Examples of static phantom types and thei r applications would be: solid phantoms imitating tissue layer density changes for needle stick training; Agar or polymer filled phantoms used for ultrasound training or calib ration; water or cont rast filled acrylic cylinders used for MRI/CT scanner calibration. The study of blood flow through organs and vasculature requires the use of “flow” phantoms [64-66] that can accommodate the “dynamic” flows and pressures of fluids such as water/contrast/g lycerol mixtures or blood. The ability to move a flow phantom between imaging modalities enables comparative assessment of each technology’s abi lity to image and quantify flow against a common physical reference. As a flow phant om must provide a geometrically accurate 3D image of the fluid flow and allow accurate measurement of fluid velocity, the desired properties of a candidate casting material are low image distortion, sharp vessel edge definition, and high contrast between the fluid and the material.

PAGE 166

145 Methods Target Application The following proposed flow phantom de sign was used to select candidate materials from the hundreds of industrial casti ng products available. The dimensions of the flow phantom were set to 10x10x10 cm (1000 cc) (1 L) (61 in3) to accommodate 3D imaging of the first three portal vein branches of a full scale Caucasian male liver. The flow phantom would be connected to inlet and outle t hoses to establish fluid flow through the phantom while the phantom rested on a lab table for ultrasound imaging, or, on the patient beds of MRI or CT systems. Desired flow phantom mate rial properties included: low cost, easily obtained, non-toxic for ease of fabrication and handling, and non-flammable in raw or cured form. Physical and dimensional stability over phys iological and imaging suite temperatures (18/26 C) (65/80 F) and physiological fl uid pressures (120 mmH g) (2.3 psi) was desired to maintain an accurate shape of the vascular cavities to ensure repeatable flow data over a six month study. A flexible material was desired to allow repeated insertion of barbed tubing connections without cracking. Easy clean ing was desired to prevent build-up of contrast/gly cerol films or blood residues. Ma terials which are chemically and biologically inert were desired so that the phantom would not degrade or decompose over time when exposed to sunlight, temperature, or the stated fluids. Optical clarity was desired for observing and resolv ing air bubble issues in the casting processes, finished casting body, and fluid flow during testing, as well as viewing of the cast anatomy for training purposes. ManufacturerÂ’s specifi cations and handling observations during

PAGE 167

146 testing were used to screen and evaluate the properties not explicitly tested in this study phase. Literature Search While usable 3D flow calibration can be achieved using materials with no anthropomorphic properties, this study focuse d on materials with densities near human tissue to acquire data that might be useful to future studies simulating human tissue properties with or without flow. Density was chosen as it was a property commonly reported in both manufacturer data and medical tissue literature. A database of tissue and casting compound properties was compile d from the references below. Materials Selection and Fabrication Several materials were selected against the desired properties for the proposed flow phantom design. Two test samples of each selected material were fabricated in accordance with the manufacturer Â’s instructions. One sample of each material was molded or drilled to create vertical and diagonal contra st wells into which imaging modality appropriate contrast materials such as wood, metal, or water could be inserted or injected. The other sample was fabricated as a solid casting with no internal cavities. Ultrasound Testing A Philips (Philips Healthcare, Andover, MA) iU22 ultrasound system was used to test the materials for use in the proposed fl ow phantom design. The samples were tested

PAGE 168

147 in a de-ionized water bath in front of a Philips V6-2 (2-6 MHz) 3D Curvilinear transducer. The transducer was protected with a thin plastic film as a water seal and acoustic gel was used under the plastic film to improve coupling. Soft wooden sticks with rough surfaces were inserted into the contrast wells of the samples to simulate the diffuse reflections obtained from powdered cornstarch which is commonly used as a contrast medium in water based ultrasound fl ow testing. Cornstarch was not used as contrast in the sample wells since it settles ou t of solution quickly, cau sing a loss of target reflectivity over time. The wooden sticks we re used to provide consistent reflection targets over long test times. Since ultrasound image quality is greatly influenced by the medium through which the sound travels from the transducer to the target and then back, several criteria were establishe d to identify usable ultrasound phantom materials. 1) The material should exhibit low attenuation as meas ured by a high pixel intensity in the target reflection, meaning that a majority of the ac oustic wave energy was able to reach the target and return. 2) The material should exhibit low backscatter as measured by a low pixel intensity in the material between the target and the transducer, meaning that the interfaces between the polymer molecules and air trapped in the material microstructure are not breaking up the coherent acoustic wave and distorting the image. The ratio of target reflection intensity to backscatter inte nsity was chosen as a relative figure-of-merit to differentiate the performance of the materials. High atte nuation and/or high backscatter would reduce image clarity, edge detection, and Doppler velocity accuracy at imaging depths where the vascular caviti es are small. The following intensity measurements were taken: front surface refl ection as a monitor of delivered acoustic

PAGE 169

148 energy, target reflection intensity, an d material backsc atter intensity. ImageJ software (http://rsb.info.nih.gov/ij/) was used to comput e the mean intensity of 200 pixels in the areas of interest. MRI Testing While NMR relaxometry would be the bett er choice for characterizing proton rich polymers, the imaging resource available was a Philips Achieva 3T (3 Tesla) MRI scanner. Each of the materials were tested for use in the proposed flow phantom design using developmental Ultrashort Echo (UTE) and standard Spin Echo (SE) pulse sequences. The contrast wells in the samples were filled with water to provide a contrast relative to the sample material. To provide the water mass required for the scanner to calibrate the hydrogen frequency, a dish of wate r was placed with the samples in a corner of the sample tray. An SE scan was perfor med to discover the contrasts of the materials relative to water and each other. As the sample materials are not uniformly hydrated with unbound water as are tissues, they were expected to have an extremely short T2*, where T2* is defined as the time constant descri bing the exponential decay of signal, due to spin-spin interactions, magnetic field inhomogene ities, and susceptibility effects. The T2* value of the samples were determined using a dual echo 3D UTE scanning sequence [67] with the following parame ters: voxel size 0.78 x 0.78 x 5 mm3; TR 45 ms; Flip Angle (FA) 15; 2392 radial trajectories; 30 slices. The achieved minimum echo time was 90 s and the following echo times were chosen for imaging: 140 s; 200 to 900 s in 100 s steps; 1000 to 10,000 s in 1000 s steps. The T2* of the samples was

PAGE 170

149 computed from the pixel signal intensity at each echo time (TE) using a 2-compartment model. According to this model, the MR signal ( y ) is given by (1) where and are the coefficient and T2* value of the short T2*compartment, respectively, and and are the coefficient and T2* value of the long T2* compartment, respectively. The term represents a Rician-distributed error term which we neglect because it is one order of magnitude smaller than max(a1,a2) for the samples. Using extrapolation for TE 0, we obtain the relation (2) and determine the value for each pixel of the samples. The unknowns of the T2* 2compartment modeling are then determin ed by solving the optimization problem (3) such that the unknowns are larger or equal to zero. The fractions of the spins for the short and long T2* compone nts are determined for each pixel by (4) (5)

PAGE 171

150 Quantitative shading plots of of the short T2* compartment and of the long T2* compartment were created for the samples. ImageJ was used to compute the mean intens ity of 200 pixels within the materials and the water reference near the center of the samples. The images of acceptable candidate materials for MRI phantoms were exp ected to show contrast wells with sharp edge detail and visible contrast between th e material and the water in the well. CT Testing A Philips Gemini 64TF PET-CT scanner calib rated for water was used to test the materials for use in the proposed flow phantom design. The contrast test wells in the samples were filled with water to provide cont rast relative to the sample material. The samples were scanned in 1 mm slices at 1 mm steps in nine combinations of exposure/dose (mAs) and tube voltage (kVp), comprising three settings of exposure for each of three tube voltage settings,. ImageJ was used to compute the mean Hounsfield Unit (HU) value of 200 pixels within the materials near the center of the sample. The images of acceptable candidate materials fo r CT phantoms were expected to show contrast wells with sharp edge detail and visible contrast between the material and the water in the well.

PAGE 172

151 Results Literature Search Table III.1 contains the phantom materi als most commonly referenced in the literature reviewed for this study. The table lists each materialÂ’s characteristics against the properties desired for the proposed flow phantom design. Table III.1 Overview of Common Phan tom Materials and Characteristics Substance Preparation Toxicity Dimensional Stability Bioactive References Agar Simple No No Yes [60, 61, 68] Aquaflex Pad N/A No No No Polyacrylamide Complex Yes No No [68, 69] Polyvinyl Alcohol Moderate Irritant No No [60, 70, 71] Silicone Simple Low Yes No [68] Polyurethane Simple Yes Yes No [72, 73] Oil Gel Simple Low No No [73] Gelatin Simple Low No Yes [61, 62] Agar is commonly used for sealed ultr asound targets or for short lived training and research models which do not require dime nsional stability. It can require additives to prevent biological growth and water im mersion and/or refrigeration to prevent decomposition in air. Aquaflex gel pads are commonly used fo r clinical ultrasound patient examinations but shrink dimensionally when exposed to air. Polyacrylamide (PAA) has toxicity issues and changes di mensionally over time. Polyvinyl alcohol (PVA) is difficult to prepare and requires sp ecial storage to prev ent decomposition. Oil Gel and Gelatin are attractive due to their avai lability and low cost but dissolve in the presence of fluid flow, are not generally se lf-supporting, and have decomposition issues.

PAGE 173

152 The literature search produced a database of human tissue an d industrial casting material properties with 200+ entries and the relevant tissues and ma terials are listed in Table III.2 sorted by the Density property. Manufacturer data sheets reported physical properties such as Hardness, Density, El ongation, and Opacity, whereas the ultrasound imaging literature (as a modality example) reported Density in addition to acoustic properties such as Velocity, Impedance a nd Attenuation. As exha ustive measurement and correlation of all properties for each material for each modality was beyond the scope of the study, the study focused on how each of the selected test materials performed against the desired imag e quality criteria. Table III.2 Properties of Human Abdo minal Tissues and Industrial Casting Materials. (The materials appearing in bold were te sted in this study. Please refer to the Nomenclature section at the end of the article for abbrev iations used in the table.) The candidate materials were selected against the desired properties of the proposed flow phantom and included five tw o-part silicones (P4, P-125, 10T, 27T, T-4) and one two-part polyurethane (RP-6400). Two-part materials involve mixing a Material Opacity Cure SG Density (kg/m3) Hard Elong % Veloc. Long. (m/s) Z (Mrayl) Attn (dB/cmMHz) Ref. Fat 924 1450 1.34 0.50 Bushberg[74] Water 1.000 998 1497 1.49 Bushberg[74] P-4 Clear CCP 1.010 1010 41 150 SI Oil Gel 1 1040 1480 0.40 Kondo[73] RP-6400 Opaque 1040 52 420 1500 1.56 Zell[68] Agar 2% 1040 1500 1.57 0.40 Zell[68] Blood 1058 1560 1.65 0.18 Bushberg[74] Oil Gel 2 1060 1580 1.80 Kondo[73] Liver 1061 1555 1.64 0.40 Johnston[75] Muscle 1068 1600 1.71 Bushberg[74] 15T CCT 1.080 1080 15 600 SO 10T Transl. CCT 1.090 1090 10 586 SO P-125 Clear CCP 1.090 1090 40 250 SI PA 1090 1580 1.73 0.70 Zell[68] PVA 1100 1570 1.74 2.90 Zell[68] T-4 Transl. AC 1100 40 400 Dow, Inc. 27T Transl. CCT 1.110 1110 27 400 Smooth-On Skin 1110 1540 1.71 9.20 Duck[76]

PAGE 174

153 base/resin with a hardener/catalyst to start th e curing process. Two of the silicones were cured by condensation with Titanium (10T, 27T), two were cured by condensation with Platinum (P-125, P-4), and one of the si licones was cured by addition (T-4). Condensation and addition curing refer to ‘two methods of vulcanizing (molecular crosslinking) liquid silicone which offer differe nces in controllability of cure rates and cured properties such as shri nkage and abrasion resistance’. The candidate materials possessed Hardness (Shore A) values from 10 (10T) to 52 (RP-6400), densities from 1010 kg/m3 (P-4) to 1110 kg/m3 (27T), and mixed viscosit ies from 7500 cps (P-4) to 100,000 cps (P-125). The costs ranged fr om $21/kg (P-4) to $50/kg (T-4). Manufacturers included: 10T & 27T from Sm ooth-On, Inc. ( Easton, PA); P-4 & P-125 from Silicones, Inc. (High Point, NC); T-4 from Dow, Inc. (Midland, MI); RP-6400 from Freeman Manufacturing and Supply (Avon, OH). Test Sample Fabrication The samples were cast in a cylindrical sh ape with a 2.54 cm (1”) diameter and a 5 cm (2”) height as shown in Fig. III.1. Post cure opacity ranged from optically clear (P-4, P-125), to translucent (10T, 27T, T-4), to opaque (RP-6400). RP-6400 exhibited an opaque golden yellow color.

PAGE 175

154 Fig. III.1 Silicone and Polyurethane Test Samples All samples were prepared according to the manufacturers specifications of a 10:1 mix of base and catalyst that typically yi elded a sample volume per unit mass of mixed material of 0.9 L/kg (25 in3/lb). Silicones P-4, P-125, T-4, 10T and 27T were prepared in a well ventilated room as recommended and did not emit noticeable noxious fumes during preparation. Polyurethane RP-6400 wa s prepared and cured under an externally vented hood as recommended. None of the samples exhibited post cure odors or surface residues. Single diagonal rectangular cavities were molded into the P-4, T-4, 10T, 27T samples. Two opposing diagonal rectangular cavities were molded into the RP-6400 sample. Two vertical round holes of differe nt diameters were drilled in the P-125 and RP-6400 samples. As the process of thoroughly mixing th e base and catalyst introduces micro bubbles, all manufacturers recommended a vacuum de-gas prior to casting. To de-gas the higher viscosity materials within working times ranging from 30 minutes (RP-6400) to 120 minutes (P-4), a thin layer of the material was placed in a flat wide shallow tray and the material was exposed to al ternating cycles of vacuum and atmospheric pressure. One manufacturer recommended cycling between a vacuum of 711 mmHg (28 inHg) and an

PAGE 176

155 air pressure of 5931 mmHg (203 inHg)(100 psi) such that the vacuum cycle would pull bubbles out of the mixture by attracting the trap ped gas, and, the air pressure would push the bubbles out by forcing the molecules in the mixture together. Depending on the material, a container volume of 5 times the vol ume of the mixture was used to allow the degassing bubbles to expand a nd “break” under vacuum. The low viscosity P-4 material de-gassed quickly within its stated worki ng time at atmospheric pressure allowing degassing to be performed in the mold itself. When degassed under vacuum, the uncured P4 was easily drawn through micr o-cracks in the mold walls. The high viscosity samples (P-125, T-4) were difficult to de-gas as bubbl es would not rise th rough any thickness of the mixture quickly (if at al l). Degassing of all materials was slowed by the curing process which increased the viscosity of th e mixture over the working time from an initially flowing mixture to a solid (fully cu red) material. It was difficult (if not impossible) to prevent re-introduction of bubbles while pouring the de-gassed mixtures into the mold and success with high viscos ity materials required post pour degassing in the mold itself. Pouring a continuous stream in one spot yielded the fewest bubbles. Ultrasound Results The ultrasound images of the test samp les are shown in Fig. III.2. The V6-2 transducer implemented a long ax is phased array scan that ap peared as a bright arc from left to right at the top of the images.

PAGE 177

156 Fig. III.2. Ultrasound Images showing Backscatter and Target Reflections The P-4 sample exhibited a bright, thin a nd uniform front surface reflection layer, low backscatter, and a bright reflection fr om the wood stick target pressed into the contrast well of the sample. The 10T sample exhibited a bright, thin and uniform front surface reflecting layer, noticeable backscatte r, and a bright reflection from the wood stick target. The 27T sample exhibited a br ight, thin and uniform front surface reflecting layer, less backscatter than the 10T, and a bright reflect ion from the wood stick target. The 10T sample exhibited more backscatter th an the harder and more dense 27T. The T4 sample exhibited a bright non-uniform fr ont surface reflecting layer, non-uniform backscatter, and a faint reflection from w ood target. The P-125 sample exhibited a bright, thin and uniform front surface reflec ting layer, no backscatter, and no reflection from the wood stick target. The RP-6400 samp le exhibited a bright, thick and very non-

PAGE 178

157 uniform reflecting layer deep into the samp le, bright widespread backscatter, and no target reflection. The ultrasound reflection measurements ar e plotted in Fig. III.3 and listed in Table III.3. The front reflection levels of a ll materials were within 10% of each other, a variation due to differences in transducer -to-sample distance, and varying material hardness (stiffness) which influe nces acoustic reflection. Fig. III.3 Ultrasound Results: Reflect ion and Backscatter Characteristics

PAGE 179

158 Table III.3 Ultrasound Results: Data for Fig. III.3 with Computed Figure-of-merit Material Front Reflection Internal Backscatter Target Reflection Figure of Merit (dB) P-4 229.9 3.5 219.3 80.0 P-125 226.9 15.8 0.0 N/A 27T 230.5 50.5 217.8 6.3 T-4 238.6 103.2 161.4 1.9 10T 249.5 108.0 221.6 3.1 RP-6400 246.3 123.9 0.0 N/A The P-4, 10T, and 27T target reflections we re slightly attenuated from the front surface reflection levels, suggesting low atte nuation despite differences in material properties and cure chemistry. The T-4 targ et reflection was notic eably lower suggesting higher attenuation and backsca tter. The P-125 target reflec tion was very low suggesting very high attenuation since the associated back scatter was very low. The RP-6400 target reflection was not measurable and further lite rature search uncovered a chemical cousin of the RP-6400, RP-6401 (Freeman), that has an attenuation of 100 dB/cm-MHz which is too high for acoustic imaging if they share this property. The P-4 backscatter was the lowest in the group and the RP-6400 was the highest. The T-4 was close to the 10T, and the 27T wa s in the middle of the range. The large difference in backscatter intensity between the 10T and 27T might be explained by more air intrusion into microstructure of the less de nse 10T and/or their difference in hardness. The P-125 backscatter was very low. The last column of Table III.3 lists the ratios of target reflection to backscatter for the materials in decibels (dB) as a figureof-merit where higher values indicate more

PAGE 180

159 reflected signal and less backsc atter. The P-4 exhibited the best performance against this metric. MRI Results Fig. III.4 shows coronal SE and UTE images of the samples in the tray with the water reference dish in the upper left corner Table III.4 lists the relative intensities between the materials and the water reference. Fig. III.4 MRI Results: Images of Spin Echo and Ultrashort Echo scans at various values of pulse repetition time (TR)(m s), echo time (TE)(ms), and Flip Angle (degrees). (The TE 0 in the figure is 140 s rounded to 0 ms)

PAGE 181

160 Table III.4. MRI Results: Material Inte nsities for SE and UTE Scan Sequences relative to water. ( The entries of less than 100 dB are a result of RP-6400 intensity values close to zero) Material SE (dB) UTE TE 140 s (dB) UTE TE 1ms (dB) RP-6400 < -100 -1.3 < -100 P-125 -22.0 -1.8 -6.3 P-4 -15.4 -2.9 -6.3 10T 0.7 -0.4 -1.6 27T -0.7 -0.4 -1.8 The left-most image in Fig. III.4 shows a T1-weighted Spin Echo scan with a TE of 10 ms, TR of 600 ms and F lip Angle of 70. The 10T and 27T samples are nearly the same intensity as the water reference dish in the upper left hand of the image, and, the 10T appears slightly lighter in intensity th an the 27T as confirmed by the ratios provided in Table III.4. The edge definition of the 10T and 27T contrast wells was distinct. The intensities of the P-4 and P-125 materials ar e indistinguishable fr om the background air intensity (black). The intens ity of the RP-6400 material wa s indistinguishable from the background air intensity (black) and the water in the contrast wells was visible with good edge definition. The middle image of Fi g. III.4 shows a UTE scan with a TE of 140 s, TR of 45 ms, and Flip Angle of 15. The 10T and 27T samples are nearly identical in intensity to the water referen ce as confirmed in Table III.4. The contrast wells of both samples are distinguishable from the material with good edge definition. The intensities of the P-4, P-125 and RP-6400 materials are nearly equal in intensity and slightly darker than the water reference. The RP-6400 cont rast wells are disti nguishable from the

PAGE 182

161 material with good edge definition. The ri ght-most image of Fig. III.4 shows a UTE scan with a TE of 1 ms, TR of 45 ms, and Flip Angle of 15. The 10T and 27T samples are slightly darker than the water reference and have similar intensities as confirmed by the data in Table III.4 and the contrast wells have good edge definition. The intensities of the P-4 and P-125 are identical a nd their contrast wells are no t visible. The RP-6400 was indistinguishable from air (black) and the contrast we lls are not visible. Fig. III.5 shows shading plots for the T2* measurements and Table III.5 lists the associated values for each material. The T 2* quantities were obtained by interpolation or extrapolation based on the data for the 2-compartment model. The UTE technique clearly differentiates silicones and polyuretha ne chemistries and shows a slight difference in the Long T2* value between the titanium cured 10T/27T and the platinum (P-4, P-125) and addition cured silicones (T-4) Fig. III.5 MRI Results: Shading pl ots of T2* material intensities

PAGE 183

162 Table III.5 MRI Results: T 2* Mean Material Intensities Material Short T2* ( s) Long T2* ( s) Short T2* (%) Long T2* (%) P-4 25 2250 5 95 P-125 25 2250 5 95 T-4 35 2250 5 95 27T 25 3300 5 95 10T 25 3300 5 95 RP-6400 535 250 95 5 CT Results Fig. III.6 provides cross sectional views of the sample CT scans detailing the contrast wells containing water. The P4 sample shows an air bubble halfway up the diagonal well, the P-125 shows an air bubble near the top of the water well, and the RP6400 shows air above both water columns. Fig. III.6. CT Results: Cross-sections showing water in contrast wells at 120 kVp/200 mAs The Hounsfield values of each material are listed in Table III.6 and plotted in Fig. III.7. The error bars in Fig. III.7 represen t the scanner measurement uncertainty as a percentage of the measured value. All of the silicone materials (P-4, P-125, T-4, 10T,

PAGE 184

163 27T) exhibited measurable and visible contrast between the water (0 HU) in the contrast wells and the material. Although the P-4 exhibi ted the lowest contrast of the silicones, it exhibited the sharpest edge detail. Table III.6 CT Results: Hounsfield values at 200 mAs Fig. III.7 CT Results: Material In tensity Variation with kVp at 200 mAs The Hounsfield values of the sample s did not vary across 100/200/250 mAs exposure settings at constant tube voltages (k Vp). The Hounsfield values of the samples did decrease non-linearly with increasing tube vo ltage as shown in Fig. III.7 at constant Material \ (HU) 80 kVp 120 kVp 140 kVp 27T 401.0 243.8 205.5 10T 371.1 219.4 182.4 P-125 364.9 211.3 175.3 T-4 352.9 212.5 171.7 P-4 264.3 133.0 101.7 RP-6400 8.1 6.8 5.3

PAGE 185

164 exposure levels. The silicones 10T and 27T tracked each other with an offset that was also visible in their relative brightness in Fig. III.6 and the measured values in Table III.6. Silicones T-4 and P-125 tracked each other very closely across the tube voltage settings despite very different cure chemistries. The polyurethane RP-6400 exhibited a nearly zero contrast to water and very low variation across all tube voltages as seen in Figs. III.6-8 and Table III.6. The physical features of the RP-6400 sample are detailed in Fig. III.8 to show that the water column above th e silicon bottom plugs (white) and below the meniscus was nearly indistinguishable from the sample material itself. Fig. 8. CT Results: Image of RP6400 at 120 kVp/200 mAs showing nearly zero contrast to water Discussion Desired Properties All of the materials were readily availa ble in 2 kg (~5 lb) quantities and were delivered within a week after ordering from the manufacturers or their local representatives. The molding cost of the proposed phantom using the P-4 silicone would

PAGE 186

165 be approximately $21, which was considered a reasonable project expense for a durable and reusable phantom. None of the materials were listed as ex treme fire hazards by the manufacturers. The fumes from liquid RP-6400 polyurethane were listed as an inhaled toxic hazard and are mitigated by use of an externally vented hood. The inner and outer dimensions of the sa mples did not change perceptibly during testing or handling. As an indirect measure of dimensional stability under flow pressure, the wood targets for ultrasound testing were slig htly larger than the contrast wells and this applied an internal st retching force on the inner diam eter of the wells. The wood targets were installed and removed many times without material loss, cracking, or change in fit or insertion/removal force. The RP-6400 was found to reproduce surface details as fine as 0.18 mm (0.007 in). All of the materials were soiled during machining and general handling and all were easily cleaned. The samples did not disso lve, change color/opacity/appearance after exposure to water or sitting in air for the several months of testing, nor did they host molds or develop odors after repeated water exposure. The optically clear samples greatly faci litated the degassing process development and the location of bubbles in water in jected into the contrast wells. Other Applications From handling the materials during testing it became evident that the differences in hardness/density properties between thes e materials and their ability to “self heal”

PAGE 187

166 small penetrations might also be exploited to create reusable needle placement training phantoms where the student would be able to f eel the changes in penetration force at the boundaries between the materials. For example, the tactile feeling of a needle penetrating skin/muscle/bone transition might be reasonabl y simulated using 10T for skin, 27T or P-4 for muscle, and P-125 or RP-6400 for bone. Phantoms containing vascular structure could give anatomically accurate positional f eedback for blood drawing or spinal tap procedure training with fluid and/or flow in the cavities of interest. The relative material contrasts documented might also be us ed to create phantoms for imaging based radiological and interventi onal procedure training. Conclusions Based on study data and the stated sel ection criteria, the P-4 silicone provided sufficient material contrast to water and e dge detail for use across all imaging modalities with the benefits of availability, low cost, dimensionally stable, non-toxic, nonflammable, durable, cleanabl e, and optical clarity. Th e RP-6400 polyurethane provided excellent edge detail in MRI testing and should be consider ed for use as a water reference in CT phantoms. The 10T and 27T silicone s should be considered for use as a water reference in MRI phantoms. Acknowledgments This work was sponsored by NSF Grant #0932339 and the University of Colorado, SOM, Department of Radiology, wh ich was greatly appreciated. CAD/CAM

PAGE 188

167 Training: Craig Lanning, Bryan Rech. Lab support: Tony Lanctot, Jennifer Wagner. Applications support: Silicones, Inc. Nomenclature Cure Cure Chemistry AC Addition Cure Attn Acoustic Attenuation CCP Condensing Cure with Platinum CCT Condensing Cure with Titanium CT Computed Tomography Imaging Elongation Dimensional st retch before tearing Hardness (Shore A) A measure of a materialÂ’s stiffness (resistance to compression) as defined by ASME standards kcps Thousands of centipoise (cps) Mix Visc. Viscosity at time of mixing Mix Life Working time befo re material cures (hardens) MRI Magnetic Resonance Imaging SG Specific Gravity Transl. Translucent Veloc. Long. Acoustic Wave Velocity in the Longitudinal Direction Z Acoustic Impedance

PAGE 189

168 References 1 G.H. Kramer, C.E. Webber, "Evalua tion of the Lawrence Livermore National Laboratory (LLNL) torso phantom by bone dens itometry and x-ray," Int.. J. Rad. Appl. Instrum. 43 795-800 (1992). 2 J.C. Blechinger, E.L. Madsen, G.R. Frank, "Tissue-mimicking gelati n-agar gels for use in magnetic resonance imaging phantoms," Med. Phys. 15 629-636 (1988). 3 J. Brunette, R. Mongrain, G. Cloutier, M. Bertrand, O.F. Bertrand, J.C. Tardif, "A novel realistic three-layer phantom for intr avascular ultrasound imaging," Int. J. Cardiovasc. Imaging 17 371-381 (2001). 4 R.O. Bude, R.S. Adler, "An easily made low-cost, tissue-like ultrasound phantom material," J. Clin. Ultrasound 23 271-273 (1995). 5 R.V. Griffith, P.N. Dean, A.L. Anderson, J.C. Fisher, "Fabrication of a tissueequivalent torso phantom for intercalibrati on of in-vivo transuranic-nuclide counting facilities," in Symposium on Advances in Radiation Protection Monitoring (Stockholm, Sweden, 1978). 6 L.K. Ryan, F.S. Foster, "Tissue equi valent vessel phantoms for intravascular ultrasound," Ultrasound Med. Biol. 23 261-273 (1997). 7 K.V. Ramnarine, T. Anderson, P.R. Hoskin s, "Construction and ge ometric stability of physiological flow rate wall-less st enosis phantoms," Ultrasound Med. Biol. 27 245250 (2001). 8 A.J. Powell, S.E. Maier, T. Chung, T. Ge va, "Phase-Velocity Cine Magnetic Resonance Imaging Measurement of Pulsatile Blood Flow in Children and Y oung Adults: In Vitro and In Vivo Validation," Pediatr. Cardiol. 21 104-110 (2000). 9 J. Rahmer, P. Brnert, J. Groen, C. Bos, "Three-dimensional radial ultrashort echo-time imaging with T2 adapted sampling," Magn. Reson. Med. 55 1075-1082 (2006). 10 K. Zell, et al., "Acoustical properties of selected tissue phantom materials for ultrasound imaging," Phys. Med. Biol. 52 N475 (2007). 11 Z. Bu-Lin, H. Bing, K. Sheng-Li, Y. Hua ng, W. Rong, L. Jia, "A polyacrylamide gel phantom for radiofrequency ablation," Int. J. Hyper. 24 568-576 (2008). 12 I. Mano, H. Goshima, M. Nambu, M. Iio, "N ew polyvinyl alcohol gel material for MRI phantoms," Magn. Reson. Med. 3 921-926 (1986). 13 A. Kharine, S. Manohar, R. Seeton, R.G. Ko lkman, R.A. Bolt, W. Steenbergen, F.F. de Mul, "Poly(vinyl alcohol) gels for us e as tissue phantoms in photoacoustic mammography," Phys. Med. Biol. 48 357-370 (2003). 14 A.R. Selfridge, "Approximate Material Proper ties in Isotropic Materials," IEEE Trans. Sonics Ultrason. 32 381-394 (1985). 15 T. Kondo, Kitatuji, M., Kanda, H., presen ted at the IEEE Ultrasonics Symposium2005 (unpublished). 16 J.T. Bushberg, The Essential Physics of Medical Imaging (2002). 17 A. Johnston, "Elements of Tissue Characterization," in Ultrasonic Tissue Characterization II, Spec. Publ. 525, edited by e. M. Linzer (National Bureau of Standards 1979). 18 F.A. Duck, Physical Properties of Tissue: A Comprehensive Reference Book (1990).

PAGE 190

169 CHAPTER IV FULL TEXT OF AIM 1 PHANTOM FABRICATION (PUBLISHED) Title The design and fabrication of two portal ve in flow phantoms by different methods Authors and Affiliations Bryan E. Yunkera)1, Gerald D. Dodd2, S. James Chen4, Samuel Chang2, Craig J. Lanning1, Ann L. Scherzinger2, Robin Shandas1, Yusheng Feng3, Kendall S. Hunter1 1 Department of Bioengineering, University of Colorado, Denver, Colorado Department of Bioengineering University of Colorado Denver/Anschutz 12700 E. 19th Ave, MS 8607 Aurora, CO 80045 Phone: 303-748-1895 Fax: 303-724-5800 E-mail: bryan.yunker@ucdenver.edu 2 Department of Radiology, Universi ty of Colorado, Denver, Colorado Department of Radiology University of Colorado SOM 12401 E. 17th Ave., Mail Stop L954 Aurora, CO 80045 Phone: (720) 848-6608 FAX: 720-848-7315 E-mail: gerald.dodd@ucdenver.edu 3 Department of Mechanical Engineering, University of Texas, San Antonio, TX University of Texas San Antonio Department of Mechanical Engineering One UTSA Circle, Mail Stop: AET 2.332 San Antonio, TX 78249-0670 Phone: 210-458-6479 FAX: 210-458-6504 Email: yusheng.feng@utsa.edu 4) Department of Medicine, Universi ty of Colorado Denver, Colorado Department of Medicine/Cardiology University of Colorado SOM 12401 E. 17th Ave., Mail Stop B132

PAGE 191

170 Aurora, CO 80045 Phone: (720) 848-6557 FAX: 720-848-7315 E-mail: james.chen@ucdenver.edu Abstract Purpose: This study outlines the design and fabr ication techniques for two portal vein flow phantoms. Methods: A materials study was performe d as a precursor to this phantom fabrication effort and the desired mate rial properties are re-s tated for continuity. A three-dimensional (3D) portal vein patte rn was created from the Visual Human database. The portal vein pattern was used to fabricate two flow phantoms by different methods with identical interior surface ge ometry using Computer Aided Design (CAD) software tools and rapid protot yping techniques. One portal flow phantom was fabricated within a solid block of clear silicone for use on a table with Ultrasound or within medical imaging systems such as MRI, CT, PET or SPECT. The other portal flow phantom was fabricated as a thin walled tubular latex st ructure for use in wate r tanks with Ultrasound imaging. Both phantoms were evaluated for usability and durability. Results: Both phantoms were fabricated successfully and passe d durability criteria for flow testing in the next project phase. Conclusions: The fabrication methods and materials employed for the study yielded durable portal vein phantoms. Keywords: Phantom, Flow, Portal, Hepatic, Imaging Purpose This study outlines the design and fabric ation of durable portal vein flow phantoms based on a study of industrial casti ng materials [77] perf ormed in a previous

PAGE 192

171 phase of the project where sm all samples of the materials were imaged with Ultrasound and in MRI and CT systems. Flow phantoms offer a range of utility, dur ability and repeatability that cannot be obtained from in-vitro organ or animal us e. The study of blood flow through the human liver is key to the study of disease and the im provement of related surgical techniques as discussed in the next three examples. 1) Ma lignant hepatic tumors can be eradicated by heating the tumor to the necrosis (cell d eath) temperature of 55 C (131 F) using a Radiofrequency (RF) probe, however, the succes s of the procedure is limited by localized blood flow that removes delivered heat [12] [78] and may prevent some or all of the tumor from reaching the point of necrosis. 2) Hepatocytes (liver cells) exposed to chemical toxins such as alcohol or diseases such as Hepatitis can die and form a fibrous scarring (fibrosis) of the liv er lobules which can increase the resistance to blood flow through the liver and result in portal hypertension [79] [80] 3) The study of hepatic metabolism [81] can also invol ve the study of blood flow. In all of these examples, the use of flow phantoms created from human [8 2] or animal imagery should provide more accurate modeling than phantoms fabricated from manufactured tubing or geometrical approximations of vessels [61, 62, 83-85]. One phantom was fabricated within a solid block of silicone with fluid flow contained within the phantom body for imagi ng with Ultrasound on a table or within medical imaging scanners such as MRI, CT, PET, and SPECT that might be damaged by uncontained liquids. As an al ternative use of the construc tion methods developed for the

PAGE 193

172 solid block phantom, another phantom was cons tructed as a thin walled tubing structure using liquid latex for use with U ltrasound imaging in water tanks. Methods and Results Portal Mesh A three-dimensional surface mesh of the portal vein was obtained from the Visual Human MaleTM (ToLTech, Inc., Aurora, CO) for use in the flow phantom fabrication as shown in the left-hand frame of Figure IV.1. Fig. IV.1 Portal Mesh in Original and Edited forms This mesh source was chosen as the software tools required to segment and extract the mesh from patient-specific imagery were not available. The portal mesh (a dataset of triangular “faces” and associated corner “ver tices”) was delivered in the standard OBJ and STL digital file format s used by CAD software. The mesh was edited with the MeshLab (www.meshlab.sourcefo rge.net) software package at the vertex/face level to remove branches that could not be practically attached to outlet ports, to close the resulting holes in the mesh, and to remove vessel fragments and unattached vertices that interfere with surface smoothing and printing.

PAGE 194

173 The “surface relaxation” function of the GeoMagic (Morissville, NC) software package was used to smooth the mesh and yi elded the optimal surface representation of the several smoothing functions offered. The smoothed portal vein mesh is shown in the right-hand frame of Fig. IV.1. Another GeoMagic function was used to reduce the number of faces and vertices used to represent the mesh to accommodate the SolidWorks 2010 (Waltham, MA) CAD software package import limit of 20,000 faces for STL based solid models. Pattern Design The cleaned and smoothed mesh was imported to the SolidWorks 2010 3D CAD software package for the addition of inlet a nd outlet pipes at the end of each vascular branch to create a casting “pattern” for the flow phantom molding process as shown in Fig. IV.2. The geometrical tube at the top of the figure is the portal vein inlet and the geometrical tubes at the botto m of the figure are the portal branch outlets. The black shaded structure between the ge ometrical tubes is the portal vein mesh converted into a solid body.

PAGE 195

174 Fig. IV.2. Solid geometry casting pattern design used for the phantom fabrications The inner diameters of the inlet and outlet ports were sized to accommodate standard plastic barbed tubing adapters [6.4 mm (1/4 in), 9.5 (3/8 in), 12.7 mm (1/2 in)] in keeping with the elongation limits of th e silicone and latex materials to prevent cracking and ripping during barb insertion and removal. To prevent the outlets from restricting the flow from the vessel ends, each vessel was connected to an outlet structure much larger than the vessel diameter. Pattern Printing The pattern design file was processe d with Stratasys (Eden Prairie, MN) Insight and Control Center software to create the files required for “printing” the pattern as a 3D plastic structure on a Stratasys Fortus 200 mc hot plastic deposition rapid prototyping printer. As the printer prints plastic at a physical resolution of 0.18 mm (0.007 in), the previous MeshLab and GeoMagic editing and smoothing of the mesh were critical steps since the printer resolution was sufficient to re produce any geometry errors in the surface.

PAGE 196

175 Two printing materials were chosen as inputs to the printing process. “WaterWorks”, a brown colored brittle material that dissolves when immersed in a base pH water solution at 65 C (150 F), was chos en for the pattern portal vein structure. ABS, a red colored durable thermoset plasti c, was chosen for supporting the horizontal structures printed with WaterWorks since the plastics cannot be prin ted into thin air. (Note that for this pattern making applicati on, the roles of these materials are reversed from the traditional printing of mechanical parts where the part is printed with the durable ABS material and the brittle di ssolvable Waterworks material is used for support. For mechanical parts, the Water Works support ma terial is dissolved post printing, leaving the desired ABS mechanical part). The 3D printing process st arts with a layer of melte d plastic deposited onto a model base that snaps into the printer tray. The plastic enters the print head from a spool of small diameter hard plastic and is then he ated to the melting point for printing and then cools to a hard plastic shortly after exiting the print head. The printi ng path of the initial layers form the base of the pattern structure and are illustrated in the left-hand pane of Fig. IV.3 where the dissolvable WaterWorks pa ttern material color is purple and the ABS support material color is teal. After prin ting each layer of melted plastic, the tray supporting the model base is lowered to allow th e next layer of hot plastic to be printed on top of the previous layer. The build up and printing path of the portal pattern design is shown in the right-hand frame of Fig. IV.3. The Stratasys Insight software allows user selection of materials layer by layer, which is convenient for creating a breakoff layer low in the base to facilitate removal of the pattern from the base after printing. The

PAGE 197

176 breakoff was implemented by depositing a laye r of the brittle Waterworks material between layers of the rigid ABS support material. More importantly, Insight allows control over the printing geometry (pitch a nd overlap of the deposited beads of hot plastic) which is critical to the creation of pattern features th at are robust enough to endure the casting and pattern dissolving pr ocess without excessive printing time and materials use (a trial and error art form). Fig. IV.3 Printing path designs for the pattern base and phantom upper structure. (Waterworks pattern material color is purpl e, ABS support material color is teal). The conical caps on the geometrical in let and outlet tubes comprise a unique invention of a self supporting roof structur e which eliminates the need for ABS support material inside of the tubes which would be neither removable nor dissolvable. The WaterWorks pattern material could have been us ed as an internal roof support structure at the expense of doubling the pr inting time which was 12 hours as implemented. The angle (slope) of the conical caps was determined by the minimum overlap of the beads of plastic deposited on top of each other without falling off the bead below, which sets the maximum horizontal bead pitch. Tubes terminated in domes or flat tops, for example,

PAGE 198

177 cannot be implemented without internal support material because the minimum overlap is always exceeded where the structure sl ope is predominantly horizontal. The fabricated portal vein pattern is shown in Fig. IV .4 with and without the ABS support material where the WaterWorks patt ern material color is brown and the ABS color is red. The smallest practical vessel diameter with this fabrication process was 1 mm (0.04 in) due to the delicat e process of physically separa ting the brittle WaterWorks pattern material from the rigid ABS support ma terial without damage to the portal vein pattern features. Fig. IV.4 Left: brown colored dissolvable Water Works pattern material printed over the red colored ABS support material. Right: finished dissolvable portal vein pattern after ABS support materi al removed by grinding tool Removing the ABS support material from under the WaterWorks pattern material turned out to be difficult and time consuming as the ABS is difficult to cut and melts into grinding tools (properties which make ABS very suitable for the printing of durable mechanical parts). In subsequent pattern fabrications, the support material was changed from ABS to the brittle WaterWorks materi al and the support stru cture printing pattern was redesigned with a very thin wall and ve ry low density inner webbing. This support structure design was easily broken away from the more robustly printed WaterWorks pattern.

PAGE 199

178 Solid Body Phantom Design A two part silicone, P-4 from Silicones, In c. (High Point, NC), was selected from the study of industrial casting ma terials [77] performed in adva nce of this project phase. P-4 exhibited many of the prope rties desired for the phantom design including: visible contrast and edge detail between the material and water, readily available, low cost, dimensionally stable, non-toxic, non-flammable, durable, cleana ble, and optically clear. Additionally, physical and dimensional stab ility was desired over physiological and imaging suite temperatures (18/26 C) (65/80 F) and at physiological fluid pressure (120 mmHg) (2.3 psi) to maintain the accurate shap e of the vascular cavities and to ensure repeatable flow data over a six month study. A flexible material was desired to allow repeated insertion of barbed tubing connec tions without cracking. Easy cleaning was desired to prevent build-up of contrast agent residues. A chemically and biologically inert material was desired so that the phantom will not degrade or decompose over time when exposed to sunlight, temperature, or in ert fluids. Optical clarity was desired for observing and resolving air bubble issues in the casting processes, finished casting body, and fluid flow during testing, as well as vi ewing the cast anatomy for training purposes. A low viscosity material was desired to faci litate the removal of bubbles introduced into the material during mixing and pouring. Ma nufacturerÂ’s specifi cations along with handling observations during material and pha ntom evaluation were used to screen and evaluate the properties not expl icitly tested in the materi als selection and fabrication study phases.

PAGE 200

179 The dimensions of the solid body flow phantom were set to 10x10x10 cm (1000 cc) (1 L) (61 in3) to accommodate 3D imaging of the full portal mesh. The cost of the solid body phantom was estimated at $23 from the target phantom volume, the material cost of $21/kg ($9.50/lb), and a yielded materi al volume per unit mass of mixed material of 0.9 L/kg (25 in3/lb), which was considered a reasona ble project expense for a durable and reusable phantom. The solid body phantom was molded in a he xagonal shape to provide flat surfaces for acoustic coupling with Ultrasound and to provide a stable surface when the phantom rests on a lab table or on im aging system patient beds. Solid Body Phantom Fabrication The solid body phantom was fabricated by placing a printed portal vascular pattern into a printed mold and pouring the silicone over the pattern. The casting process followed techniques developed in the materi als study [77] to minimize entrapment of bubbles within the phantom that would create undesirable imaging artifacts. The process steps of the phantom casting process are s hown in Fig. IV.5. The left-hand photo shows the pattern installed in the hexagonal phantom mold which was also printed using the WaterWorks pattern material. The middle photo shows the disso lvable pattern cast within the silicone block prior to the process of dissolving th e pattern in a 65 C (150 F) jet bath of base pH water. The right-most photo shows the fully de-molded solid body phantom and the internal void in the shape of the portal vein bran ching vasculature.

PAGE 201

180 Fig. IV.5 “Solid Body” Silicone Phantom Casting Process Steps After 12 hours in the dissolving solution, the pattern was not dissolving at an acceptable rate. The rate was increased by conn ecting the portal inlet of the phantom to the solution pump outlet which forced di ssolving solution thro ugh the phantom under pressure. Even with this modification, the solution preferred to fl ow through the larger and more open vessels than through smaller vessels. The flow through the pattern was inhibited by the low porosity of the pattern design, an issue corrected in the thin walled phantom design and fabrication. A slight “fog” can be seen in the silic one casting of the solid body phantom in the right-hand frame of Fig. IV.5 that is cau sed by dissolving solution forced into the polymer microstructure under pressure duri ng the pattern dissolvi ng process. The fog disappeared after several hours in open air. A slight optical blurring can be seen on the hexagonal sides of the phantom in Fig. IV.5 (middle) and Fig. IV.6 which is an optical effect caused by the slight surface roughne ss of the mold at the 0.18 mm (0.007 in) printer step resolution. This surface smoothness effect would be diminished with use of a printer with higher resolution.

PAGE 202

181 Solid Body Phantom Evaluation The solid body phantom is shown in Fig. IV.6 on a lab bench with barbed inlet and outlet fittings plumbed to a gravity fed wa ter source for leak and durability testing. Fig. IV.6 “Solid Body” Phantom plumbed for leak testing The barbed fittings were easily installed and there were no leaks around the inlet/outlet seals or from the phantom body itself over long periods containing standing fluid. The silicon microstructure did not visibly abso rb water under gravity driven leak test pressures and flows. The silicon microstruc ture did not trap a cornstarch contrast dissolved in the test water flow. A water/ glycerol mix was not used to simulate blood over concern that the more viscous glycerol would not come out of the phantom cavities and would create a permanent echogenic cl oud around the fluid channels that would degrade future Ultrasound images and Doppl er measurements. Blood was not used for leak testing for concern that it would pe netrate the microstructure and decompose, preventing the phantom from being cleaned to a sterile state. Further smoothing and sealing of the vessels with a lacquer coating of the interior of the portal vascular voids

PAGE 203

182 was not implemented for concern the coating would create an impedance mismatch or reflecting surface for Ultrasound at the vessel walls The optical clarity of the P-4 silicon greatly facilitated air bubble removal from th e leak test flow streams. The elongation characteristic of the P-4 silicone was sufficiently high to allow good seals around the barbed tubing fittings without leaking and prevented cracking of the phantom block if attention was paid to the fitting insertion for ce and the use of a small amount of silicone grease. In future versions, half of each barb fitting will be cast into the phantom itself to provide full mechanical strength for tubing support. Thin Walled Phantom Construction As a precursor to the thin walled phant om construction, an Ultrasound evaluation of commercially available thin walled (0 .4 0.8 mm) (1/64 – 1/32 in) latex, white silicone, and TygonTM tubing samples was performed to guide the material selection. The latex samples exhibited the lowest front su rface reflection and best lumen definition. The setup for testing commercial dimensional tubi ng is shown in Fig. IV.7 with latex test specimens installed. Fig. IV.7 Ultrasound test appara tus for commercial di mensional tubing

PAGE 204

183 A liquid latex (Liquid Latex Mold Maki ng Rubber, AeroMarine, Inc., San Diego, CA) was selected for molding the irregular shap e of the portal structure. The thin walled phantom was fabricated by dipping the printed hepatic vascular pattern into the liquid latex. The fabricated thin walled phantom is shown in the left-hand frame of Fig. IV.8 and was double dipped to ensure coverage ye t maintain an average wall thickness of under 1 mm (0.04 in). Fig. IV.8 “Thin Walled” Phantom and installation in water tank The latex had a low surface tension, a high vi scosity, and a long drying time that required rotation of the dipped pattern sl owly in three dimensions whil e the liquid latex dried. The rotation process caused some surfaces on the complex pattern to retain more material than others resulting in minor thickness differences. The fabrication included the formation of a flat round base of latex whic h maintained the alignment of the otherwise flexible and floppy outlet legs and facilitated barb inserti on and seal integrity. The pattern dissolving process was accelerated by re-designing the printing paths of the pattern to create a hollow core throughout the phantom pattern structure. This feature allowed the entire phantom to dissolve from th e inside out instead of from end to end as with the solid body phantom. In future, a fu rther reduction in the dissolving time will be achieved by creating a distribution manifold for the solution pump so that solution can be

PAGE 205

184 forced through each outlet separately to ensure pressure and flow though each vessel using the portal vein inlet as the drain. The middle frame of Fig. IV.8 shows the barbed hose f ittings holding the phantom securely to the other side of a thic k polyethylene sheet that was used to mount the phantom in the ultrasound water tank and prov ide a barrier to wate r circulation in the tank as shown in the right-hand frame of Fig. IV.8. Thin Walled Phantom Evaluation The thin walled phantom was plumbed to a gravity fed water source for leak and durability testing under static and dyna mic flow conditions well above normal physiology. The barbed fittings sealed with no leaks and there were no ruptures in the phantom walls. When immersed in the wate r tank the phantom stay ed rigid between the static water pressure outside the phantom and the static and dynamic pressure of the fluid flow within the phantom. The latex materi al in the thickness fabricated was very resistant to tearing and cutting suggesting that phantoms with even thinner walls would be usable with improvements in dipping/soli d bodying technique. Th e color of the latex darkened over time with exposure to air a nd water but the material showed no signs of physical degradation over several months. On the path to improved thickness uniformity, the authors have performed preliminary tests with a “spray-on” latex which may enable phantoms of “membrane” thickness.

PAGE 206

185 Conclusions Solid body silicone and thin walled latex flow phantoms can be successfully fabricated using the commercially availa ble CAD software tools and rapid prototyping methods presented. Acknowledgments This work was sponsored by: Colorado Tr anslational Research Imaging Center (C-TRIC); NSF 0932339; University of Colo rado, SOM, Department of Radiology; NIH T32HL072738; K25 HL094749; K24081506; RO1HL114753; and NHLBI K25-094749 which was greatly appreciated. Lab s upport: Bryan Rech, Tony Lanctot, Jennifer Wagner. Applications suppor t: Silicones, Inc., SmoothOn, Inc.. Hepatic models: ToLTech, Inc. Appendix: Nomenclature ABS A rigid thermoplastic (a crylonitrile-butad iene-styrene) CAD Computer Aided Design CT Computed Tomography MRI Magnetic Resonance Imaging PET Positron Emission Tomography SPECT Single Photon Emission Computed Tomography Elongation The ability of a material to be stretched without tearing

PAGE 207

186 References a) Corresponding author: bryan.yunker@ucdenver.edu 40 209 21 28 27 11 32 17 23 32 26

PAGE 208

187 CHAPTER V CONCLUSIONS Based on study data and the stated sel ection criteria, the P-4 silicone provided sufficient material contrast to water and e dge detail for use across all imaging modalities with the benefits of availability, low cost, dimensionally stable, non-toxic, nonflammable, durable, cleanabl e, and optical clarity. Th e RP-6400 polyurethane provided excellent edge detail in MRI testing and should be consider ed for use as a water reference in CT phantoms. The 10T and 27T silicone s should be considered for use as a water reference in MRI phantoms. Solid body silicone and thin walled late x flow phantoms can be successfully fabricated using the commercially availabl e CAD software tools and rapid prototyping methods presented. While the solid sili cone portal phantom was not usable for ultrasound imaging due to high acoustic at tenuation, the thin walled latex phantom provided excellent Color Doppler lumen image definition and Pulse Wave velocity measurement capability. The CFD simulation and measured flows compared within 17% over the normal hepatic pressure range indicat ing that CFD modeling can be used to predict the flow characteristics of anatomical phantoms from their mesh construction. The volumetric flows computed from measuring the mass outle t flows of the net flow phantom and from Pulse Wave Doppler angle corrected velocitie s compared within +/ -5%, indicating that

PAGE 209

188 either can be used to calculate volume and mass flow rates if the tubing or vessel diameter is known. As designed, both ADP and VAS allowed reli able user selection of arbitrary ROIs over hundreds of captures during development and testing and provided real-time user manipulation of the ultrasound da taset views and functions. The testing for Aims 3-6 was performe d with known oversized Color Doppler data volumes that allowed functional and rep eatability verificati on of the algorithms but not accuracy determinations. The VAS segmentation algorithm work ed reliably over the hundreds of segmentations performed during developmen t and testing performed. The current centerline algorithm is sensit ive to surface roughness in Color Doppler and requires a smooth Doppler data volume as input. Th e segmentation and centerline algorithms worked well with the ~2x ove rsized Doppler data volumes. The diameter algorithm acquired diamet ers from the oversized Doppler data volumes successfully; however, the estimates di d not compare favorably to the phantom tubing inner diameter since the Doppler cross-sections were kn own to be ~2x larger than the tubing lumens due to blooming. The deviat ion in the random diameter measurements across for non-tapered images was < 9%, s uggesting that the algorithm will work well with sufficiently echogenic fluids. The diam eter algorithm is susceptible to errors in centerline direction and the number of voxe ls captured using the Dot Product method. While the diameter algorithm can accommodate some amount of vessel eccentricity and

PAGE 210

189 truncation, a more complex algorithm is need ed to accommodate strongly elliptical crosssections. As designed, the angle correction algor ithm successfully located the transducer origins from the edges of high intensity BW 3D ultrasound data volumes but was not able to do so from images without a full 2D enve lope. Under near laminar flow conditions, the estimated centerline and mean angle corr ected velocities from the random samples were within 15% of the ultrasound cart values respectively, with deviations under 35%. The angle correction algorithm is sensitive to both errors in centerline position and to turbulence in the flow that may change th e velocity local to the sample point. The net flow measured through an ROI captured from the net flow phantom tubing was estimated near net zero (1.7 cc/s) in a two tube example and within 15% over 10 samples from straight and curved tubes de spite use of oversized Doppler data. The modular software architecture approach taken with the clinical application was very stable and allowed new features to be successfully added and tested with minimal effort. With close attention paid to the forms of data stored and the data movement for computation and display, the C/C++/OpenGL/Qt platform is capable of meeting the target Time-to-Results re quirement. The methods developed for characterizing flow in this study are highly depe ndent on: 1) the availability of data and calibration constants exported from the ul trasound cart in propri etary formats; 2) sufficient Color Doppler sensitivity to obtai n an adequate surface for reconstruction and segmentation; 3) successful clinician adjustme nt of the Color Doppler within the vessel lumen. To accommodate these issues and fo r the convenience of use in the Operating

PAGE 211

190 Room, the process of segmentation, automa ted angle correction, and flow calculation would be better implemented within an ultr asound cart with full control over the signal acquisition and processing. Final validation of the methods described wi ll require access to a stable contrast source. The clinical application (VAS) went beyond the capabilities of QLab to meet the clinical requirements and TTR, however, VAS is a prototype level effort in comparison to QLab, which is a mature commercial product.

PAGE 212

191 CHAPTER VI FUTURE RESEARCH There are a number of tasks to be comple ted to prepare VAS for clinical trials. The testing performed in Chapter II must be re done with blood as a test fluid to assess whether Color Doppler can be used successfu lly. The proposed test environment is a small water tank with a single thin walle d tube of known and accurate dimensions suspended within the tank for full view ultras ound scanning of the tubeÂ’s length. A small peristaltic pump will be used to circulate a small volume (1 unit) of out of date donor blood through the submerged tube at five ve locities and pressures within the normal hepatic range. A precision valve with a fine ad justment capability will be used to provide normal hepatic backpressures and flow rates in conjunction with the pump settings. The test setup will be designed to ensure parabol ic flow profiles can be achieved. Color and Power Doppler images will be captured at each velocity/pressure level to ensure that the Doppler data volumes are smooth and within th e vessel lumens. If luminal Doppler data can be obtained, VAS will be used to estimate vessel diameters and flows and the estimation algorithms will be optimized to achie ve the best practical accuracy possible. An elliptical diameter estimation algorithm is proposed for accommodation of noncircular vessels in conjunction with an RO I capture modification that will allow voxels slightly outside the ROI to be considered in the diameter calculation. The cross-sectional voxel capture algorithm may be improved by vary ing the capture tolera nce with the angle to the Cartesian axes. A new algorithm will be developed to automate the collection and

PAGE 213

192 plotting of velocity profiles. Influence of surface noise on the cente rline may be reduced by implementation of a surface smoothing algor ithm to be run prior to the center-lining algorithm, and/or, implementation of a splin e fit approximation after centerlines are computed. In summary, all of the error source s discussed in Chapter II must be revisited with the intent of eliminating or minimi zing their contributions to acceptable and repeatable levels. The user interface needs to be simplified for ease and speed in clinical use including fewer and larger c ontrol buttons and the implem entation of a drag-to-select track ball and touch screen capability. The enti re application must be analyzed (profiled) to assess the run-time speed of each function for opportunities to consolidate functionality and optimize performance. The application functions and architecture must be reviewed against newer releases of OpenGL and Qt for opportunities to add functionality or improve speed and response time. While flow estimation is helpful to the planning process, the ultimate goal is a clinical application that can be used for ab lation planning, which requires the addition of ablation estimation capabilities based on heat models and va lidation with bovine based RF ablation testing prior to human trials. Although a PC laptop based version of the c linical application would be sufficient for animal and clinical trials, the application is currently locked to one cart manufacturer as 3D DICOM implementations and private tags are generally proprie tary for ultrasound. Implementing the clinical application on the ultrasound cart itself would greatly improve the commercial viability and clinical use of the application as the cart is used for

PAGE 214

193 monitoring ablation progress and needle location and it is the source for the data needed by the clinical application. A clinician requested feature is the inte gration of RF probe position information to assist the positioning of the RF probe with the target ablation zone and tumor. As shown in previous chapters, VAS displays ultr asound data in the display panel such that the clinician can view the scene from any pers pective and proximity to the individual data points. In the a broader context of the OpenGL language, the scene viewed by the clinician is a “virtual world” in which the ultrasound data volumes and graphical lines and dots are actually individual OpenGL “objects” which can be added or subtracted to the scene as well as rotated, translated and s caled independently. Af ter detection, the RF probe can be represented in th e 3D scene as a probe shaped object. As a demonstration of this capability for use in future re search work, Figs. VI.1 and VI.2 show the “C/OpenGL” version of VAS displaying the po rtal and hepatic vein “objects” extracted from the Visual Human along with other “obj ects” in the form of walls, an operating table, and a patient, to cr eate a “virtual” surgery.

PAGE 215

194 Figure VI.1 Extracted Vasculature Registered to Patient Physiology The figure show the vascular structures extracted from the visual human images overlaid and registered with transparent a nd non-transparent renderings of the images they were extracted from to give a represen tation of the segmented anatomy relative to the patient’s entire body. The scene is cen tered inside a cube which has photographs applied to its interior surfaces as textures. The operating table consists of two cuboids objects with photographs applied as textures to their exterior surfaces. The patient object includes the 180 visual human image slices wh ich can be shown as a transparency (Fig. VI.1) or as opaque images (Fig. VI.2). This capability could be implemented for training, pre-operative planning, or real time guided surgery. The clinician is able to rotate the view of the patient and zoom in to and literally “fly through” the room and through the patient anatomy using keyboard and mouse inputs.

PAGE 216

195 Figure VI.2 Virtual Reality Demo Follow-on Capabilities Although 3D DICOM file formats are propr ietary, there are many available 3D (multi-slice) data sets for MRI and CT which are not, opening an opportunity for a version of VAS which uses the BW boundaries of tissue/blood/airway and tissues as the segmentation boundary reference. The study of vessel and/or airway structure over time might yield informative trends in pediatri c growth, disease progression and even early disease detection. Further, VAS could be enhanced to co-register images from all modalities to aide diagnostics and possibly prediction. As the VAS architecture currently acc ommodates surfaces displayed as point clouds, the next natural step is to create a representation of that surface in the “face and vertex” format used by many CA D design and graphics sub-syst ems. Representation of the vessel surfaces as faces would enhance the fly-through visual experience by blocking

PAGE 217

196 the view to graphical features behind the su rface, and, enable the use of OpenGL lighting (already used in coursework) and pipeline shader capabilities to give added clinical detail to the structure. A prototype surfaci ng algorithm was attempted in the C/OpenGL version of VAS and can be vaguely seen in Fig. VI.1 as red hepatic vein vasculature. It is not illustrated or discussed in further deta il as it needs more work to perfect. Other features can be added to VAS with a significant amount of research into existing methods. As many images result in “spatially noisy” or “physically rough” surfaces, it would be appropriate to add su rface smoothing capability to VAS similar to the “relaxation” techniques used in Geomag ic. The current centerline finding algorithm implemented for VAS does not accommodate vessel branches. The most promising branching centerline algorithm found to da te is the Laplacian reduction [86] which requires a manifold surface to iteratively reduce to zero diameter, however, from initial inspection of the algorithm, it would be accurate but not meet the clinical TTR. As the surface of the segmented structures represen t a “pattern dependent” mesh as commonly used in commercial CAD analysis software, it would be possible to add functions to VAS which would create a volumetric sub-mesh (t etrahedral, etc.) and th en add functions to perform Finite Element and Finite Volume computations for flow and heat transfer analysis for more accurate (albeit non-TTR) es timations. This would be a very attractive real-time capability in contrast to use of commercial analysis packages designed to handle millions of points and very large structures which are very expensive and time consuming to set up.

PAGE 218

197 REFERENCES 1. Samuel Chang, P.R., Anthony Lanctot, Liping Huang, Martin, McCarter, Gerald Dodd. Comparison of 3D Volumetric Color D oppler Vascular Indices to Predict the Performance of Radiofrequency Ablati on Devices in Liver Tumors in In-vitro Blood Perfused Bovine Liver Model in ARRS annual meeting 2012. Vancouver, Canada. 2. American Cancer Society Occurance 2014; Available from: http://www.cancer.org/cancer /livercancer/detailedguide/ liver-cancer-what-is-keystatistics. 3. American Cancer Society Survival Rates Available from: http://www.cancer.org/cancer /livercancer/detailedguide /liver-cancer-survivalrates. 4. Cancer Research UK Worldwide Statistics Available from: http://www.cancerresearchuk.org/cancerinfo/cancerstats/world/mortality/#Common. 5. Wong, S.L., et al., American Society of Clinical Oncology 2009 clinical evidence review on radiofrequency ablation of hepa tic metastases from colorectal cancer. J Clin Oncol, 2010. 28 (3): p. 493-508. 6. Siegel, R., D. Naishadham, and A. Jemal, Cancer statistics, 2013. CA: A Cancer Journal for Clinicians, 2013. 63 (1): p. 11-30. 7. Lu, D.S.K., et al., Radiofrequency Ablation of Hepatocellular Carcinoma: Treatment Success as Defined by Histologi c Examination of the Explanted Liver1. Radiology, 2005. 234 (3): p. 954-960. 8. Yin XY, X.X., Lu MD, et al. Percutaneous thermal ablation of medium and large hepatocellular carcinoma: lo ng-term outcome and prognostic factors. Cancer, 2009. 115 (9): p. 1914-23. 9. Dodd, G.D., 3rd, et al., Minimally invasive treatm ent of malignant hepatic tumors: at the threshold of a major breakthrough. Radiographics, 2000. 20 (1): p. 9-27. 10. Dodd, G.D., 3rd, et al., Radiofrequency thermal ablatio n: computer analysis of the size of the thermal injury created by overlapping ablations. AJR Am J Roentgenol, 2001. 177 (4): p. 777-82. 11. Choi D, L.H., Rhim H, et al., Percutaneous radiofrequency ablation for earlystage Hepatocellular carcinoma as a first -line treatment: long-term results and prognostic factors in a larg e single-institution series. Eur Radiol, 2007(17): p. 684-692. 12. Montgomery, R.S., A. Rahal, and G.D. Dodd, Radiofrequency ablation of hepatic tumors: variability of lesion size using a single ablation device. AJR Am J Roentgenol, 2004(182): p. 657-661. 13. Chinn SB, L.F.J., Kennedy GD, et al., Effect of vascular occlusion on radiofrequency ablation of the liv er: results in a porcine model. AJR, 2001(176): p. 789-795.

PAGE 219

198 14. Welp C, S.S., Ermert H, et al., Investigation of the influ ence of blood flow rate on large vessel cooling in hepatic radiofrequency ablation. ; :. Biomed Tech (Berl), 2006(51): p. 337-46. 15. Chinn, S.B., et al., Effect of vascular occlusion on radiofrequency ablation of the liver: results in a porcine model. AJR Am J Roentgenol, 2001. 176 (3): p. 789-95. 16. Lu, D.S., et al., Effect of vessel size on creation of hepatic radiofrequency lesions in pigs: assessment of the "heat sink" effect. AJR Am J Roentgenol, 2002. 178 (1): p. 47-51. 17. Germer, C.T., et al., Experimental study of l aparoscopic laser-induced thermotherapy for liver tumours. Br J Surg, 1997. 84 (3): p. 317-20. 18. Shen, P., et al., Laparoscopic radiofrequency ablati on of the liver in proximity to major vascularture: Effect of the Pringle maneuver. J Surg Oncol, 2003. 83 : p. p. 36. 19. de Baere, T., et al., Percutaneous radiofrequency ablation of hepatic tumors during temporary venous occlusion. AJR Am J Roentgenol, 2002. 178 (1): p. 53-9. 20. Patterson, E.J., et al., Radiofrequency ablation of porci ne liver in vivo: effects of blood flow and treatmen t time on lesion size. Ann Surg, 1998. 227 (4): p. 559-65. 21. Chang, C.K., et al., Radiofrequency ablation of the porcine liver with complete hepatic vascular occlusion. Ann Surg Oncol, 2002. 9 (6): p. 594-8. 22. S. Rossi, F.G. and I.D. Francesco, Relationship between the shape and size of radiofrequency induced thermal le sions and hepatic vascularization. Tumori, 1999. 85 : p. p. 128. 23. Wacker, F.K., et al., MR-guided interstitial laserinduced thermotherapy of hepatic metastasis combined with arte rial blood flow reduction: technique and first clinical results in an open MR system. J Magn Reson Imaging, 2001. 13 (1): p. 31-6. 24. Bitsch, R.G.M., et al., Effects of Vascular Perfus ion on Coagulation Size in Radiofrequency Ablation of Ex Vivo Perfused Bovine Livers. Investigative Radiology, 2006. 41(4) (April ): p. 422-427. 25. Huang, L., G.D. Dodd, 3rd, and A.C. Lanctot, Radiofrequency ablation of the liver: effect of variation of portal venous blood flow on le sion size in an in-vitro perfused bovine liver. Acad Radiol, 2012. 19 (8): p. 1018-22. 26. Moriyasu, F., et al., "Congestion index" of the portal vein. American Journal of Roentgenology, 1986. 146 (4): p. 735-739. 27. Carlisle, K.M., et al., Estimation of total hepatic blood flow by duplex ultrasound. Gut, 1992. 33 (1): p. 92-7. 28. Burroughs, A.K., The Hepatic Artery, Portal Venous System and Portal Hypertension: The Hepatic Veins and Liver in Circulatory Failure in Sherlock's Diseases of the Liver and Biliary System 2011, Wiley-Blackwell: London. p. 152-209. 29. Rokni-Yazdi, H., Assessment of Normal Doppler Parameters of Portal Vein and Hepatic Artery in 37 Healthy Iranian Volunteers. Iranian Journal of Radiology, 2006. 4 (3): p. 213-216.

PAGE 220

199 30. Cosar, S., et al., Doppler and gray-scale ultrasou nd evaluation of morphological and hemodynamic changes in liver va scualture in alcoholic patients. European Journal of Radiology, 2005. 54 (3): p. 393-399. 31. Tasu, J.P., et al., Hepatic Venous Pressure Gradients Measured by Duplex Ultrasound. Clinical Radiology, 2002. 57 (8): p. 746-752. 32. Brown, H.S., et al., Measurement of normal portal venous blood flow by Doppler ultrasound. Gut., 1989. 30 (4): p. 503-9. 33. Weinreb, J., et al., Portal vein measurements by real-time sonography. American Journal of Roentgenology, 1982. 139 (3): p. 497-499. 34. Ulusan, S., T. Yakar, and Z. Koc, Evaluation of portal venous velocity with Doppler ultrasound in patients with nonalcoholic fatty liver disease. Korean Journal of Radiology, 2011. 12 (4): p. 450-5. 35. Armonis, A., D. Patch, and A. Burroughs, Hepatic venous pressure measurement: an old test as a new prognostic marker in cirrhosis? Hepatology, 1997. 25 (1): p. 245-8. 36. Kumar, A., P. Sharma, and S.K. Sarin, Hepatic venous pressure gradient measurement: time to learn! Indian J Gastroenterol, 2008. 27 (2): p. 74-80. 37. Feldschuh, J. and Y. Enson, Prediction of the normal blood volume. Relation of blood volume to body habitus. Circulation, 1977. 56 (4 Pt 1): p. 605-12. 38. Levy, M., Cardiovascular Physiology 9 ed. 2007. 39. Jensen, J.A., Estimation of Blood Velocities Using Ultrasound: A Signal Processing Approach 2013, Denmark: Technical University of Denmark. 40. Bushberg, J., The Essential Physics of Medical Imaging 2012: Wolters Klewer. 41. Kreith, F.B., W., Basic Heat Transfer 1980: Harper&Row. 42. Kirbas, C. and F. Quek, A review of vessel extracti on techniques and algorithms. ACM Comput. Surv., 2004. 36 (2): p. 81-121. 43. Frericks, B., et al., 3D CT modeling of hepatic ve ssel architecture and volume calculation in living donat ed liver transplantation. European Radiology, 2004. 14 (2): p. 326-333. 44. Kaften, A Two-Stage Approach for Fully Au tomatic Segmentation of Venous Vascular Structures in Liver CT Images. 45. Wink, O., W.J. Niessen, and M.A. Viergever, Multiscale vessel tracking. Medical Imaging, IEEE Transactions on, 2004. 23 (1): p. 130-133. 46. Noble, J.A. and D. Boukerroui, Ultrasound image segmentation: a survey. Medical Imaging, IEEE Transactions on, 2006. 25 (8): p. 987-1010. 47. Guerrero, J., et al., Real-time vessel segmentation and tracking for ultrasound imaging applications. IEEE Trans Med Imaging, 2007. 26 (8): p. 1079-90. 48. Kawajiri, S., et al., Automated segmentation of hepa tic vessels in non-contrast Xray CT images. Radiol Phys Technol, 2008. 1 (2): p. 214-22. 49. Mirhassani, S.M., M.M. Hosseini, and A. Behrad, Improvement of Hessian based vessel segmentation using two stage threshold and morphological image recovering in Proceedings of the 6th international conference on Innovations in information technology 2009, IEEE Press: AI-Ain, United Arab Emirates. p. 96100.

PAGE 221

200 50. Chapman, An Analysis of Vessel Enhancem ent Filters Based on the Hessian Matrix for Intracranial MRA. 51. Erdt, M., M. Raspe, and M. Suehling, Automatic Hepatic Vessel Segmentation Using Graphics Hardware in Proceedings of the 4th international workshop on Medical Imaging and Augmented Reality 2008, Springer-Verlag: Tokyo, Japan. p. 403-412. 52. Frangi, A., et al., Multiscale vessel enhancement filtering in Medical Image Computing and Computer-Assisted Interventation — MICCAI’98 W. Wells, A. Colchester, and S. Delp, Editors. 1998, Springer Berlin / Heidelberg. p. 130-137. 53. Flasque, N., et al., Acquisition, segmentation and tracking of the cerebral vascular tree on 3D magnetic resonance angiography images. Medical Image Analysis, 2001. 5 (3): p. 173-183. 54. Agam, G., S.G. Armato, 3rd, and C. Wu, Vessel tree reconstruction in thoracic CT scans with application to nodule detection. IEEE Trans Med Imaging, 2005. 24 (4): p. 486-99. 55. Lorenz, C., et al., Multi-scale line segmentation w ith automatic estimation of width, contrast and tangential direct ion in 2D and 3D medical images in CVRMed-MRCAS'97 J. Troccaz, E. Grimson, and R. Msges, Editors. 1997, Springer Berlin / Heidelberg. 56. Esneault, S., C. Lafon, and J.L. Dillenseger, Liver vessels segmentation using a hybrid geometrical moments/graph cuts method. IEEE Trans Biomed Eng, 2010. 57 (2): p. 276-83. 57. Yunker, B.E., et al., An investigation of industrial molding compounds for use in 3D ultrasound, MRI, and CT imaging phantoms. Med Phys, 2013. 40 (5): p. 052905. 58. Yunker, B.E., et al., The design and fabrication of two portal vein flow phantoms by different methods. Medical Physics, 2014. 41 (2): p. 023701. 59. Kramer, G.H. and C.E. Webber, Evaluation of the Lawren ce Livermore National Laboratory (LLNL) torso phantom by bone densitometry and x-ray. International Journal of Radiation Applica tions and Instrumentation, 1992. 43 (6): p. 795-800. 60. Blechinger, J.C., E.L. Madsen, and G.R. Frank, Tissue-mimicking gelatin-agar gels for use in magnetic resonance imaging phantoms. Medical Physics, 1988. 15 (4): p. 629-36. 61. Brunette, J., et al., A novel realistic th ree-layer phantom for intravascular ultrasound imaging. International journal of cardiovascular imaging, 2001. 17 (5): p. 371-81. 62. Bude, R.O. and R.S. Adler, An easily made, low-co st, tissue-like ultrasound phantom material. Journal of Clinical Ultrasound, 1995. 23 (4): p. 271-3. 63. Griffith, R.V., et al., Fabrication of a tissue-eq uivalent torso phantom for intercalibration of in-vivo trans uranic-nuclide counting facilities in Symposium on Advances in Radiation Protection Monitoring 1978: Stockholm, Sweden. 64. Ryan, L.K. and F.S. Foster, Tissue equivalent vessel phantoms for intravascular ultrasound. Ultrasound in Medicine & Biology, 1997. 23 (2): p. 261-73.

PAGE 222

201 65. Ramnarine, K.V., T. Anderson, and P.R. Hoskins, Construction and geometric stability of physiological flow rate wall-less stenosis phantoms. Ultrasound in Medicine & Biology, 2001. 27 (2): p. 245-250. 66. Powell, A.J., et al., Phase-Velocity Cine Magnetic Resonance Imaging Measurement of Pulsatile Blood Flow in Children and Young Adults: In Vitro and In Vivo Validation. Pediatric Cardiology, 2000. 21 (2): p. 104-110. 67. Rahmer, J., et al., Three-dimensional radial ultrashort echo-time imaging with T2 adapted sampling. Magnetic Resonance in Medicine, 2006. 55 (5): p. 1075-1082. 68. Zell, K. and et al., Acoustical properties of select ed tissue phantom materials for ultrasound imaging. Physics in Medicine and Biology, 2007. 52 (20): p. N475. 69. Bu-Lin, Z., et al., A polyacrylamide gel phantom for radiofrequency ablation. International Journal of Hyperthermia, 2008. 24 (7): p. 568-76. 70. Mano, I., et al., New polyvinyl alcohol gel ma terial for MRI phantoms. Magnetic Resonance in Medicine, 1986. 3 (6): p. 921-6. 71. Kharine, A., et al., Poly(vinyl alcohol) gels fo r use as tissue phantoms in photoacoustic mammography. Physics in Medicine and Biology, 2003. 48 (3): p. 357-70. 72. Selfridge, A.R., Approximate Material Propert ies in Isotropic Materials. IEEE Transactions on Sonics and Ultrasonics, 1985. 32 (3): p. 381-394 73. Kondo, T., Kitatuji, M., Kanda, H. New tissue mimicking materials for ultrasound phantoms in IEEE Ultrasonics Symposium 2005. Rotterdam, Netherlands. 74. Bushberg, J.T., The Essential Physics of Medical Imaging 2002. 75. Johnston, A., ed. Elements of Tissue Characterization Ultrasonic Tissue Characterization II, Spec. Publ. 525, ed. e. M. Linzer. 1979, National Bureau of Standards 76. Duck, F.A., Physical Properties of Tissue: A Comprehensive Reference Book Academic. 1990, London. 77. Yunker, B.E., et al., An investigation of industrial molding compounds for use in 3D ultrasound, MRI, and CT imaging phantoms. Medical Physics, 2013. 40 (5): p. 9. 78. Goldberg, S.N., et al., Radio-frequency tissu e ablation: effect of pharmacologic modulation of blood flow on coagulation diameter. Radiology, 1998. 209 (3): p. 761-767. 79. Rockey, D.C., Hepatic blood flow regulation by stellate cells in normal and injured liver. Seminars in Liver Disease, 2001. 21 (3): p. 337-49. 80. Sarin, S.K., K.K. Sethi, and R. Nanda, Measurement and correlation of wedged hepatic, intrahepatic, intrasplenic and in travariceal pressure s in patients with cirrhosis of liver and non-ci rrhotic portal fibrosis. Gut, 1987. 28 (3): p. 260-266. 81. Barbaro, B., et al., Doppler sonographic assessment of functional response of the right and left portal venous branches to a meal. Journal of Clin ical Ultrasound, 1999. 27 (2): p. 75-80. 82. Taylor, C.A. and C.A. Figueroa, Patient-specific modeling of cardiovascular mechanics. Annu Rev Biomed Eng, 2009. 11 : p. 109-34.

PAGE 223

202 83. Raine-Fenning, N.J., et al., Determining the relationship between threedimensional power Doppler data and true blood flow characteristics: an in-vitro flow phantom experiment. Ultrasound in Obstetri cs and Gynecology, 2008. 32 (4): p. 540-550. 84. Inglis, S., et al., An anthropomorphic tissue-mi micking phantom of the oesophagus for endoscopic ultrasound. Ultrasound in Medicine & Biology, 2006. 32 (2): p. 249-259. 85. Taylor, C.A., T.J. Hughes, and C.K. Zarins, Finite element modeling of threedimensional pulsatile flow in the abdominal aorta: relevance to atherosclerosis. Ann Biomed Eng, 1998. 26 (6): p. 975-87. 86. Chowriappa, A., et al., 3-d vascular skeleton ex traction and decomposition. IEEE J Biomed Health Inform, 2014. 18 (1): p. 139-47.