dilated residual u-net for multi-organ segmentation in ...€¦ · dilated residual u-net for...
TRANSCRIPT
Dilated Residual U-Net for Multi-organ Segmentation in
Thoracic CT
Is 2D better than 3D segmentation?
Sulaiman Vesal1, Nishant Ravikumar1, Andreas Maier1
1Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU)
ISBI2019 SegTHOR Challenge
2April 4th - 2019Sulaiman Vesal | ISBI2019 SegTHOR | Dilated Residual U-Net for Multi-organ Segmentation in Thoracic CT
Radiotherapy & Organs at Risk (OAR)
Figure: Risks of side effects in the surrounding organs increases during breast cancer radiotherapy.
Radiation
Radiation
Cancer
Heart exposed to
radiation during
radiotherapy of left
sided breast cancer
3April 4th - 2019Sulaiman Vesal | ISBI2019 SegTHOR | Dilated Residual U-Net for Multi-organ Segmentation in Thoracic CT
Motivation
Axial View Sagittal View Coronal View
4April 4th - 2019Sulaiman Vesal | ISBI2019 SegTHOR | Dilated Residual U-Net for Multi-organ Segmentation in Thoracic CT
Motivation
Axial View Sagittal View Coronal View
Esophagus
Trachea
Heart
Aorta
5April 4th - 2019Sulaiman Vesal | ISBI2019 SegTHOR | Dilated Residual U-Net for Multi-organ Segmentation in Thoracic CT
3D U-Net[1]
[1] Çiçek et. al. MICCAI 2016
[2] Yee et. al. 2017
6April 4th - 2019Sulaiman Vesal | ISBI2019 SegTHOR | Dilated Residual U-Net for Multi-organ Segmentation in Thoracic CT
3D U-Net[1]
[1] Çiçek et. al. MICCAI 2016
[2] Yee et. al. 2017
3D U-Net Output LGE-MRI Volume Ground Truth
7April 4th - 2019Sulaiman Vesal | ISBI2019 SegTHOR | Dilated Residual U-Net for Multi-organ Segmentation in Thoracic CT
Pre-ProcessingPre-Processing
Figure: Pre-processing pipeline for SegTHOR Data
Cropping
(Nx288x288)(Nx512x512)
CLAHE
112 px
112 px
112 px
112 px
8April 4th - 2019Sulaiman Vesal | ISBI2019 SegTHOR | Dilated Residual U-Net for Multi-organ Segmentation in Thoracic CT
2D UNet + DR Architecture
9April 4th - 2019Sulaiman Vesal | ISBI2019 SegTHOR | Dilated Residual U-Net for Multi-organ Segmentation in Thoracic CT
2D UNet + DR Architecture
10April 4th - 2019Sulaiman Vesal | ISBI2019 SegTHOR | Dilated Residual U-Net for Multi-organ Segmentation in Thoracic CT
Loss Functions
● Multi-class Dice Loss[1]:
No of classes Predicted voxel of class k Ground truth
[1] Milletari et. al. MICCAI 2016
[2] Salehi et. al. MICCAI MLMI 2017
● Tversky Loss[2]:● Tversky Loss[2]:
11April 4th - 2019Sulaiman Vesal | ISBI2019 SegTHOR | Dilated Residual U-Net for Multi-organ Segmentation in Thoracic CT
SegTHOR Dataset[1]
● 60 CTs from patients with non small cell lung cancer
● Resolution: 0.98×0.98×2.50mm3
● Dimensions:
● N×512×512 voxels
● N → 150 to 284 slices
[1] https://competitions.codalab.org/competitions/21012#learn_the_details-overview
● Data Split:
40 Patients 20 Patients
Training data Testing data
5 Fold-Cross Validation
12April 4th - 2019Sulaiman Vesal | ISBI2019 SegTHOR | Dilated Residual U-Net for Multi-organ Segmentation in Thoracic CT
ISBI SegTHOR
DSC: 0.921(0.016)
HD: 0.016(0.174)DSC: 0.942(0.028)HD: 0.092(0.054)
DSC: 0.909(0.062)
HD: 0.304(0.161)
Methods Training DSC Validation DSC Testing DSC Testing HD
2D DR U-Net 0.978 0.936 0.887 0.446
2D DR U-Net (Augmented) 0.974 0.945 0.914 0.254
2D DR U-Net (Augmented) + Tversky Loss 0.975 0.946 0.916 0.250
Patient_41 Patient_44Patient_55
13April 4th - 2019Sulaiman Vesal | ISBI2019 SegTHOR | Dilated Residual U-Net for Multi-organ Segmentation in Thoracic CT
Quantitative Analysis
Figure: Violin box plot of Dice score for individual organs for 2D UNet + DR method
14April 4th - 2019Sulaiman Vesal | ISBI2019 SegTHOR | Dilated Residual U-Net for Multi-organ Segmentation in Thoracic CT
Quantitative Analysis
Figure: Violin box plot of Hausdorf distance for individual organs for 2D UNet + DR method
15April 4th - 2019Sulaiman Vesal | ISBI2019 SegTHOR | Dilated Residual U-Net for Multi-organ Segmentation in Thoracic CT
ISBI SegTHOR Leaderboard
16April 4th - 2019Sulaiman Vesal | ISBI2019 SegTHOR | Dilated Residual U-Net for Multi-organ Segmentation in Thoracic CT
Conclusion
● 2D U-Net+DR outperformed 3D U-Net with less trainable parameters
● Incorporating global contextual information improve the performance
● Fast (≈ 5.08 seconds) and robust method
● MR images, may be beneficial for some OARs with poorly-visible boundaries such as the esophagus
17April 4th - 2019Sulaiman Vesal | ISBI2019 SegTHOR | Dilated Residual U-Net for Multi-organ Segmentation in Thoracic CT
Thank you for your attention!
https://www5.cs.fau.de/~vesal
Questions?