Many medical image analysis techniques require an initial localization and segmentation of anatomical structures. As part of the VISCERAL benchmarks on Anatomy segmentation, a hierarchical multi-atlas multi-structure segmentation approach guided by anatomical correlations is proposed. The method begins with a global alignment of the volumes and refines the alignment of the structures locally. The alignment of the bigger structures is used as reference for the smaller and harder to segment structures. The method is evaluated in the ISBI VISCERAL testset on ten anatomical structures in both contrast-enhanced and non-enhanced computed tomography scans. The proposed method obtained the highest DICE overlap score for some structures like kidneys and gallbladder. Similar segmentation accuracies compared to the highest results of the other methods proposed in the challenge are obtained for most of the other structures segmented with the method.
Anatomical correlations for a hierarchical multi-atlas segmentation of the pancreas in CT images
1. Anatomical correlations for a hierarchical
multi-atlas segmentation of the pancreas in
CT images
Oscar A. Jiménez del Toro
University of Applied Sciences
Western Switzerland
(HES-SO)
4. Introduction
• Anatomical segmentation is fundamental for
further image analysis1
• Different methods proposed2,3 (regression
random forests, level set…)
• Comparison of multiple approaches for the
same public dataset is uncommon
4
5. VISual Concept Extraction challenge in
RAdioLogy
• EU funded project (2012-2015)
– HES-SO, ETHZ, UHD, MUW, TUW,
Gencat
• Organize competitions on medical
image analysis on big data
• All computation done in the cloud
• Segmentation benchmark
• Retrieval benchmark
• Annotation by medical doctors
7. Benchmark 2 Anatomy
• Automatic segmentation of
anatomical structures (20)
and landmark detection
• Define challenges in large
scale data (aprox. 10TB)
processing
• CT and MR images
(contrast-enhanced and
non-enhanced)
10. Hierarchical multi-atlas segmentation
• Use multiple atlases for
the estimation on a target
image
• Global and local alignment
• Hierarchical selection of
the registrations improves
results
• Label fusion
17. Right
Kidney
Liver
Global
alignment
Urinary Bladder Right Lung Left Lung1st Lumbar Vertebra
Gall-
bladder
Left KidneyTrachea
Spleen
2nd Local
Affine
Affine
Local Affine
B-spline non-
rigid
Liver
Right
Kidney
Pancreas segmentation
18. Experimental setup
• VISCERAL Benchmark 1 testset
• 10 contrast-enhanced CT volumes of the trunk
• Added to segmentation method of 10
structures:
– Liver, lungs, kidneys, gallbladder, urinary bladder,
1st lumbar vertebra, trachea and spleen
• 7 independent atlases as trainingset
19. Results
• Average DICE score for Pancreas: 0.52
Structure DICE Rank in VISCERAL
Benchmark 1
Liver 0.918 1st
Right Kidney 0.913 1st
Left Kidney 0.921 1st
Right Lung 0.965 3rd
Left Lung 0.955 3rd
Spleen 0.852 3rd
Trachea 0.836 2nd
Gallbladder 0.566 1st
Urinary bladder 0.7 3rd
1st Lumbar vertebra 0.522 2nd
20. Results
• Average DICE score for Pancreas: 0.52
Structure DICE Rank in VISCERAL
Benchmark 1
Liver 0.918 1st
Right Kidney 0.913 1st
Left Kidney 0.921 1st
Right Lung 0.965 3rd
Left Lung 0.955 3rd
Spleen 0.852 3rd
Trachea 0.836 2nd
Gallbladder 0.566 1st
Urinary bladder 0.7 3rd
1st Lumbar vertebra 0.522 2nd
21.
22. Conclusion
• Full automatic method
• Requires little or no feedback from the user
• Showed robustness in the segmentation of
multiple structures with high overlap
• Fared well when compared to other methods
of the VISCERAL Benchmark 1
• Future work:
– Extend to method to other modalities (CTwb ISBI challenge, MR)
– Test in a bigger dataset for VISCERAL Benchmark 2 Anatomy