The document describes the Liver Workbench project, which developed an image-based liver workbench with 3D segmentation, modeling, and quantification toolkits for clinical applications like tumor volumetry, characterization, and surgical planning. Key features include robust liver segmentation from CT scans, a CADx system for detecting and characterizing liver lesions, and an interactive surgery planning system to design resections while ensuring safety margins. The workbench aims to support clinical decision-making and research.
1. Liver Workbench −
An Integrated Tool-Suite for Liver Components Segmentation,
Quantification and Surgical Planning from CT Data
A project funded by the Joint Council Office (JCO), A*STAR
ZHOU Jiayin*, CHEN Wenyu, HUANG Weimin, XIONG Wei, Thiha OO
Institute for Infocomm Research (I2R), A*STAR, Singapore
LIU Jimin, CHI Yanling, TIAN Qi
Singapore Bio-imaging Consortium (SBIC), A*STAR, Singapore
Sudhakar K. VENKATESH
National University Hospital (NUH), Singapore
Liver Workbench
11/2012
Contact: 6408-2497, jzhou@i2r.a-star.edu.sg
2. Motivation
Liver cancer: serious threaten to human health with 0.6-1.0 M
new cases per year
Surgical resection / transplantation offers the best prognosis
Precise liver surgery expands the availability of liver surgery
Surgery planning has increasing demands for quantitative
analysis of liver components
Gross liver, liver segments, tumors, vascular structure……
Objective
Construct a liver CT image database with associated ground truth
for benchmarking and building statistical models
Develop a Liver Workbench with 3D liver object segmentation,
modeling and quantification toolkits
Clinical applications: tumor volumetry, tumor characterization and
surgical planning
Liver Workbench
11/2012
3. Liver Workbench (An image-based liver workbench with 3D liver object segmentation,
modeling and quantification toolkits for clinical applications)
Oct 2009 ~ Apr 2013, a JCO funded project collaborating with SBIC and NUHS
Surgical resection / transplantation offers the best prognosis for liver cancer treatment.
Surgery planning has increasing demands for quantitative analysis of liver structures.
A Liver Workbench with 3D liver object segmentation, modeling and quantification
toolkits is being developed to explore various of clinical applications.
Project Architecture
Liver 3D object segmentation
(Liver, tumor, vessel, etc)
Liver 3D object quantification,
validation & modeling
Liver 3D model interaction &
visualization
Probabilistic
Atlas
CT/MRI
Database
Clinical applications
3D liver/tumor
volumetry
Liver Workbench
11/2012
Tumor type
characterization
Pre-operative
planning
More….
4. Modules / Technologies Developed
3D Liver & Liver Tumor Segmentation
3D Liver Vasculature Extraction
Modeling: Construction of Probabilistic Liver Atlas
Focal Liver Lesion Detection & Characterization
Surgical Planning for Transplant and Tumor Removal
Important Features
1. A robust platform to segment and quantify liver and its component
from CT scans;
2. A CADx system to detect and characterize focal liver lesions;
3. An intuitive and flexible way to plan liver surgery interactively;
4. Support clinical decision-making and biomedical research.
Liver Workbench
11/2012
5. 3D Liver Segmentation (WACV 09’, RSNA 09’)
•
3D Liver Volume Segmentation by Flipping-free Mesh Deformation
and Registration
Uses explicit quadrilateral mesh representation and Laplacian deformation
for the purpose of efficiency;
Solves self-intersection problem by detecting and discarding possible
flippings on mesh surface before each iteration;
Incorporates shape constraints to reduce sensitivity to noise;
Easy to implement
Liver Workbench
11/2012
6. 3D Liver Segmentation
Test on clinical CT volume - liver segmentation
20 sets of CT-scan data, with slice thickness from 1-3 mm
Compared with level-set and 2D grab-cut.
Min.
Max.
Mean
STD.
Median
Relative average volume difference (RAVD, %)
0.0
30.8
7.1
8.7
3.5
Volumetric overlap error (VOE, %)
6.6
36.3
12.3
7.1
9.9
Average symmetric surface distance (ASSD, mm)
1.1
10.5
2.5
2.1
1.8
W/o flip avoidance
Liver Workbench
11/2012
The dynamic evolution procedure
With flip avoidance
7. Liver Tumor Segmentation (MICCAI-MLMI 11’, EMBC 13’)
•
Liver Tumor Segmentation by Hybrid Support Vector Machine
(SVM) Classifier
Combination of the advantages of one class SVM and binary SVM
Automatic generation of balanced training data
Results from one single study
Working steps
Liver Workbench
11/2012
8. Liver Tumor Segmentation
Test on clinical CT volume - liver tumor segmentation
15 sets of CT-scan data with 26 tumors, with slice thickness from 1-3 mm
13 for parameters tuning and 13 for test
Overall segmentation of the liver, liver tumor and gallbladder
Liver Workbench
11/2012
9. Liver Vessel Segmentation (IEEE-TBME 11’)
•
Liver Vessel Segmentation by Vessel Context-based Voting
The liver has an unique dual blood supply
system – Hepatic artery, portal vein and
hepatic vein
Hepatic vascular structure determines the
partitioning of liver segments
Surgical planning requires accurate
analysis of vascular structure
Touching
Vessels
By level set
Proposed
Over-segmented
Under-segmented
Seperated
Vessels
Working steps
Liver Workbench
11/2012
10. Liver Structure Modeling (ICIP 09’, RSNA 09’)
• Construction of A Probabilistic Liver Atlas
An pair of atlases encoding probabilities of liver anatomic
and structure variabilities
An atlas retaining densitometric mean
An atlas retaining spatial variance
Helps segmentation, interpretation, group comparison, etc
Key task: To register images from different subjects to a
common coordinate system
The proposed landmark-free registration method:
Registration based on dense correspondence of all voxels without landmarks
Multiple dataset registration is unbiased to all datasets registered
Registration is in infinite dimensional diffeomorphic space
Probabilistic analysis in both density and geometry
Tested using 30 CT scans, 5 mm section thickness
Liver Workbench
11/2012
11. Liver Structure Modeling
Registration convergence
of
mean square errors
Unbiased registered multi-organs
M SE
14000
MSE5
M S E 10
M S E 15
M S E 20
M S E 25
12000
10000
8000
6000
4000
2000
Anterior view
Posterior view
Unbiased registered liver
0
1
2
3
4
5
6
Iterations
7
8
9
10
1 iteration
5 iterations
10 iterations
The mean images (gray) and respective probabilistic atlases (red)
Liver Workbench
11/2012
12. Liver Lesion Detection & Characterization (SPIE 11’, RSNA 11’)
Patent filed
Arterial
Portal vein
Delayed
Visual detection of small-size focal liver
lesions (FLLs) can be difficult;
Characterizing FLLs is usually
experience-dependent;
Detect focal liver lesion by subtracting
normal liver parenchyma and vessels
from liver region.
Characterize focal liver lesion using similarity retrieval based on multiple phase
CT image features
Creation of database using 87 confirmed cases with 6 types
Leave-one-out for testing using multiple parameters
Texture feature and its derivatives
Density feature and its derivatives
Easy retrieval of lesions with different pathology but similar appearances
Retrieval of lesions with same pathology but different appearances
Assist in decision-making on radiological diagnosis by providing evidence
Train medical students and radiological residents
Liver Workbench
11/2012
13. Liver Lesion Detection & Characterization (IJCARS 13’, Med Phys 13’)
IJCARS 13’
Medical
Physics 13’
Liver Workbench
11/2012
14. Interface
Similar cases
Query
3D View
#2 #1
NC
#2
#1
ART
PV
Two big tumors are detected.
Top 1 candidate: 104 ml and
83% similar to a confirmed
FNH.
Top 2 candidate: 155 ml and
88% similar to a confirmed
cyst.
DL
Retrieval results
Top 1
Top 2
Top 3
Top 4
Top 5
Top 6
Top 7
Top 8
Load Query
Preprocessing
FLL detection
Top 9
Top 10
Top 11
Top 12
Top 13
Top 14
Top 15
Top 16
FLL retrieval
Reporting
15. Liver Surgery Planning (RSNA 12’, EMBC 13’, MICCAI-MIAR 13’)
•
An Interactive Liver Surgery Planning System
Comprehensive real-time 3D visualization and mesh deformation
Plan, design and adjust the resection map with graft/remnant volumetry
Automatic guarantee of the safety margin with the minimal resection surface
The Main User Interface
Volumes of lobes and
the percentages
Liver Workbench
11/2012
Planning of hemi-hepatectomy with MHV preservation
16. Liver Surgery Planning
Adjust the Resection Surface to MHV Harvesting
Left and right
lobes with PV
Update the volume change
Liver Workbench
11/2012
18. Liver Surgery Planning
Liver, vasculature and tumor are segmented from CT data and
the 3D graphical model is created.
Show 10 mm tumor margin (red sphere)
Anterior-superior view
posterior-superior view
Only show hepatic vein (HV)
Only show portal vein (PV)
19. Liver Surgery Planning
A rough hepatectomy resection
plane, with the constraint to 10
mm tumor margin
Liver Workbench
11/2012
A more precise resection surface,
with the constraint to 10 mm
tumor margin
23. Summaries
1. A robust platform to segment and quantify liver and its component
from CT scans;
2. A CADx system to detect and characterize focal liver lesions;
3. An intuitive and flexible way to plan liver surgery interactively;
4. Support clinical decision-making and biomedical research / drug
development.
Segmentation of a liver with its components
Liver Workbench
11/2012
Cum Laude
Award
RSNA 12’
Liver surgical planning