Registration of 3D Ultrasound to Computed Tomography Images of the Kidney Jing Xiang Supervisors: Dr. Robert Rohling Dr. Purang Abolmaesumi Electrical and Computer Engineering University of British Columbia 1
Registration of 3D Ultrasound to Computed Tomography Images of the
Kidney
Jing Xiang
Supervisors: Dr. Robert Rohling
Dr. Purang Abolmaesumi
Electrical and Computer Engineering University of British Columbia
1
2
Motivation
• Amongst Canadians, kidney cancer is the 6th most common cancer in women and 10th most common cancer in men (Canadian Urological Association).
• Incidence rate of kidney cancer has increased by approximately 1.3% per year for males and females since late 1990s (Canadian Cancer Statistics 2010).
• For diagnosis, patients receive CT angiography (CTA).
• Partial nephrectomy is preferred over radical nephrectomy.
Anatomy of the Kidney
3
Overview
4
• Provide a suitable validation platform for registration. This entails developing a soft tissue phantom that can create realistic images in both US and CT and provide a gold standard for alignment.
• Develop a fully automatic registration method to align US and CT images of kidney using rigid-body registration.
• Examine two approaches to US simulation-based registration that differ in how the ultrasound simulation is created.
• Acquire clinical data and demonstrate how the registration performs with the additional challenges that arise with real subjects.
5
Thesis Objectives
Acquisition of Phantom Data
US: Sonix RP machine with 3-7 MHz convex curvilinear abdominal probe (4DC7-3)
CT: Aquilion 64-slice CT scanner (Toshiba Medical Systems)
6
7
Goals of Phantom Construction
• High quality images in CTA and US.
• Depict the surface boundaries of the kidney.
• Define the vascular and pyramid anatomy of the kidney in both modalities.
8
Comparison
A) US of non-freshly excised porcine kidney with no contrast.
B) CT of non-freshly excised porcine kidney with no contrast.
A
C) US of freshly excised porcine kidney with contrast.
C
B
D
D) CTA of freshly porcine excised kidney with contrast
Phantom Construction
9
A) Remove the renal capsule so that it does not trap air.
B) Inject contrast agent
(Omnipaque iohexol):
• 1 to 40 dilution in water to highlight the cortex.
• 1 to 5 dilution in gelatin solution to highlight the arteries.
A
B
10
Phantom Construction Cont’d
C) Artery and vein are separated.
D) Artery and vein are tied off to prevent leaking of contrast into the agar.
C
D
11
Phantom Construction Cont’d
• A) Kidneys are positioned.
• B) Agar is poured over and cooled to set.
• Can identify the vascular system, renal pyramids and the renal cortex.
12
Comparison to Human CTA
A B
Acquisition of Patient Data
US: Sonix RP machine with 3-7 MHz convex curvilinear abdominal probe (4DC7-3)
CT: Somatom Sensation 64-slice CT scanner (Siemens Medical Systems)
13
• Patient is placed in flank position with a cushion beneath the abdomen.
• The preoperative US and CT are taken with the patient mimicking the position during surgery.
14
Acquisition of Patient Data
Registration Methods
15
Register
16
Simulation of US from CT
Replace with:
Wein et al., 2008
Air
z1
z2
17
Simulated US
18
Pre-scan and Post-scan Converted US
Scan Conversion
Pre-scan Converted Data
Post-scan Converted Data
19
Methods of Simulation
20
Rigid Registration Algorithm
CMA-ES Optimizer Hansen et al., 2003
• Initial alignment obtained by finding the fiducial markers in both CT and US using Horn’s method (Horn, 1987) to determine the six rigid registration parameters.
21
Kidney Phantom Experiment Details
• For each test, the CT was perturbed by a transform where each parameter was selected from a range of ±10° rotation about each axis and ±10 mm translation along each axis. • 50 registration tests were run. • 7 phantoms in total.
22
Results: Kidney Phantoms
• Method A: mean TRE ranged from 1.8 to 3.9 mm. • Method B: mean TRE ranged from 1.4 to 4.2 mm.
23
Renal Cancer Patient Experiment Details
• Bronze standard alignment was obtained using Principal Components Analysis (PCA) on manually segmented surfaces of the US and CT.
• For the registration tests, a set of transforms was selected randomly with different misalignment errors.
• 20 registration tests were performed for each range of misalignment errors: 0 – 5 mm, 5 – 10 mm, 10 – 15 mm and 15 – 20 mm.
24
Initial Alignment of Patient Dataset
25
Results: Renal Cancer Patient
26
Results: Renal Cancer Patient
• Developed a detailed recipe for constructing soft tissue phantoms to evaluate CTA to US registration on kidneys.
• Validated simulation-based registration with the CMA-ES optimizer on phantom data using two different simulation approaches.
• Tested simulation-based registration on patient data and initial results demonstrate an improvement in registration accuracy by using a directionally simulated US.
27
Conclusion: Contributions
• Validation on Patient Data. – Finding a method of acquiring a gold standard alignment
between US and CT.
– Extensively validate registration on more clinical data sets.
• GPU acceleration of the registration method. – Implement with parallel processing such that registration can
be close to real-time.
• Investigate need for deformable registration.
• Compare simulation-based registration to other approaches such as feature-based approaches.
28
Conclusion: Future Work
• Funding NSERC
• Supervision Dr. Robert Rohling and Dr. Purang Abolmaesumi
• Clinical Support Dr. Chris Nguan Vickie Lessoway and Chris Eddy
• Data Acquisition Jack Bell Research Center Canada Diagnostics Centers Vancouver General Hospital
Acknowledgements
29
30
Extra Slides
31
Effect of Stand-off Pad on Image Quality
32
Fiducials in Patient Data
33
PCA
• Segmented surfaces are aligned using the major and minor axes • Concavity of the surface is used to determine the correct orientation.
34
P-curve
35
Formulation of the Simulated US
36
Scan Conversion
37
Scan Conversion Correction
38
Scan Conversion Correction
39
Scan Conversion Correction
Slow Half of the Sweep Fast Half of the Sweep
40
Understanding Simulation in the Direction of the US Beam
• For multimodality problems, a suitable similarity metric is mutual information (MI).
• However, according to Wein et al., 2008, there are too many possible configurations where the joint entropy is minimal.
• At the correct alignment, there is only a small local optimum.
41
Mutual Information?
42
Stages of Simulation for Phantom Data
A
B
C
D
E
• Khamene et al, 2006
43
Comparison of Similarity Metrics for Portal Imaging to CT Registration