Date post: | 04-Jun-2015 |
Category: |
Health & Medicine |
Upload: | university-of-applied-sciences-western-switzerland |
View: | 255 times |
Download: | 1 times |
Evaluation of a Hierarchical Anatomical Segmentation Approach in VISCERAL Anatomy Benchmarks
Oscar Jiménez-del-Toro Henning Müller University of Applied Sciences Western Switzerland (HES-SO)
2
• Motivation • VISCERAL
• Method
• Experimental Setup • Results Anatomy 1 Benchmark
• Discussion
• Conclusions
Overview
3
• Motivation • VISCERAL
• Method
• Experimental Setup • Results Anatomy 1 Benchmark
• Discussion
• Conclusions
Overview
• Anatomical segmentation is fundamental for further image analysis and Computer-Aided Diagnosis1
• Manual annotation and visual inspection is time consuming for radiologists
• Accurate large scale data analysis techniques are needed
4
Motivation
5
• Motivation • VISCERAL
• Method
• Experimental Setup • Results Anatomy 1 Benchmark
• Discussion
• Conclusions
Overview
6
• Motivation • VISCERAL • Method
• Experimental Setup • Results Anatomy 1 Benchmark
• Discussion
• Conclusions
Overview
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 o Anatomy benchmarks
o Detection benchmark
o Retrieval benchmark
VISCERAL Anatomy Benchmarks • All computations done in the
cloud
• Annotation by medical doctors
• Automatic segmentation of anatomical structures (20) and landmark detection
• CT and MR images (contrast-enhanced and non-enhanced)
9
• Motivation • VISCERAL
• Method
• Experimental Setup • Results Anatomy 1 Benchmark
• Discussion
• Conclusions
Overview
10
• Motivation • VISCERAL
• Method • Experimental Setup • Results Anatomy 1 Benchmark
• Discussion
• Conclusions
Overview
Hierarchic Multi Atlas-Based Segmentation2 • Use multiple atlases for label estimation • Global and local alignment • Hierarchical selection of the registrations
• Reuse registrations from the bigger structures (eg. liver) for the smaller ones
• Label fusion
• Atlas = Patient volume + labels • Coordinate transformation
increases spatial correlation between images3
• Multi-scale gaussian pyramid4
Image registration
Affine alignment
• Global rigid align
• Independent local refinement for bigger structures (eg. liver, lungs)
• Regions of interest based on the morphologically dilated initial estimations
• Non-rigid b-spline
• Multi-scale approach • Faster optimization
due to better initial alignment
Non-rigid alignment
Label fusion
• Majority voting threshold • Classification on a per-voxel
basis • Local registration errors are
reduced • Threshold optimization
Right Kidney
Liver Urinary Bladder
Right Lung
Left Lung
1st Lumbar Vertebra
Gall- bladder
Left Kidney Trachea
Spleen 2nd Local
Affine
Hierarchical Registration
Affine
Local Affine
B-spline non-rigid Label fusion
Global alignment
17
• Motivation • VISCERAL
• Method
• Experimental Setup • Results Anatomy 1 Benchmark
• Discussion
• Conclusions
Overview
18
• Motivation • VISCERAL
• Method
• Experimental Setup • Results Anatomy 1 Benchmark
• Discussion
• Conclusions
Overview
Experimental setup
• VISCERAL Anatomy 1 • Testset :12 contrast-enhanced CT of the trunk
• Applied to 10 anatomical structures
• VISCERAL Anatomy 2 • Testset :10 contrast-enhanced CT of the trunk
• 10 unenhanced whole body CT
• Applied to ALL anatomical structures
20
• Motivation • VISCERAL
• Method
• Experimental Setup • Results Anatomy 1 Benchmark
• Discussion
• Conclusions
Overview
21
• Motivation • VISCERAL
• Method
• Experimental Setup • Results Anatomy 1 Benchmark • Discussion
• Conclusions
Overview
Anatomy 1 Results DICE coefficient
DICE coefficient
Top rank in benchmark
Anatomy 1 Results
Anatomy 1 Results Average distance error
Anatomy 1 Results Average distance error
Top rank in benchmark
Discussion
Comparison with other participant methods: - SJ: Spanier et al. Rule-based approach with
region growing for multiple organs - HJ: Huang et al. Multiple prior knowledge
models and free-form deformation - W: Wang et al. Fast model bases level set
method and hierarchical shape priors - K: Kazmig et al. Clustering and graph cut
using shortest path constraint for spatial relations
- GG: Gass et al. Multiple atlases via Markov random field registrations
(DICE)
Discussion
Comparison with other participant methods: - SJ: Spanier et al. Rule-based approach with
region growing for multiple organs - HJ: Huang et al. Multiple prior knowledge
models and free-form deformation - W: Wang et al. Fast model bases level set
method and hierarchical shape priors - K: Kazmig et al. Clustering and graph cut
using shortest path constraint for spatial relations
- GG: Gass et al. Multiple atlases via Markov random field registrations
(DICE)
Discussion • Competitive results compared with up to 5
segmentation methods in Anatomy1
• Similar to state-of-the-art methods for some organs: liver (0.89-0.96)5,6, kidneys (0.92-0.98)7,8
• Segments not only abdominal organs but can be implemented for any anatomical structure
• Future work: Extend to method to other modalities
Conclusion • Straightforward automatic multi-structure
segmentation method
• Showed robustness in multiple structures particularly for ceCT
• High overlap for the bigger structures (e.g. liver, lungs) and competitive overlap for smaller structures (e.g. gallbladder)
Thank you for your attention !!
References 1 K.Doi. Current status and future potential of computer-aided diagnosis in medical imaging. British Journal of Radiology, 78:3-19, 2005 2Jiménez del Toro et al., Multi-structure Atlas-Based Segmentation using Anatomical Regions of Interest. Proceedings of Medical Image Computing and Computer Assisted Intervention (MICCAI2013) MCV workshop, Nagoya, Japan, 2013 3Stefan Klein et al. Elastix: a toolbox for intensity-based medical image registration. IEEE Transactions on medical imaging, 29(1):196-205, 2010 4 Stefan Klein et al. Adaptive stochastic gradient descent optimisation for image registration. International Journal of Computer Vision, 81(3):227-239, 2009
References 5Criminisi et al. Regression forests for efficient anatomy detection and localization in computed tomography scans. Medical Image Analysis, 17(8):1293-1303, 2013 6Okada et al. Abdominal multi-organ segmentation of CT images based on hierarchical spatial modeling of organ interrelations. Abdominal Imaging 2011, 7029:173-180, 2012 7Zhou et al. Automatic localization of solid organs on 3D CT images by a collaborative majority voting decision based on ensemble learning. Computerized Medical Imaging and Graphics, 36:304-313, 2012 8Wolz et al. Multi-organ abdominal CT segmentation using hierarchically weighted subject-specific atlases. Proceedings of Medical Image Computing and Computer Assisted Intervention (MICCAI2012), 7510:10-17, 2012