Liver segmentation using U-net: Practical issues @ SNU-TF
- 1. Liver segmentation using U-net; Practical issues Wonjoong Cheon / Ph.D. candidate Department of health Science and Technology, SAIHST, Sungkyunkwan University, Seoul 06351, Korea Liver Segmentation
- 2. 1. Data preparation 1-1. dataset 1-2. Preprocessing 2. Architecture 2-1. U-net Application 2-2. Information of in-house U-net 3. Result 3-1. Results: good & bad (1)~(4) 3-2. Discussion 4. Paper review
- 3. 1.1 Dataset Dtype: *.nii Modality: Computed tomography (CT) Data (train): 24 subjects(3812) Data (test) : 3 subjects(649) Dimension: 512 × 512 × z axis Dtype: *.jpg Modality: Computed tomography (CT) Data (train): 11 subjects(2664) Data (test) : 4 subjects(351) Dimension: 512 × 512 × z axis LiTS (Liver Tumor Segmentation Challenge) Liver imaging (Some person draw ROI)
- 4. 1.2 Preprocessing Standardization (Z-score scaling) • 자료 집합에 적용되는 전처리 과정으로 모든 자료에 다음과 같은 선형 변환을 적용하여 전체 자료의 분포를 평균 0, 분산 1이 되도록 만드는 과정이다. • 스케일링은 자료의 overflow나 underflow를 방지하고 독립 변수의 공분산 행렬의 condition number 를 감소시켜 최적화 과정에서의 안정성 및 수렴 속도를 향상시킨다. 추후, Batch normalization 으로 변경. 정량적으로 약 2 %의 Accuracy 향상을 보임. (https://www.youtube.com/watch?v=TDx8iZHwFtM, PR-12) Image Resize Resize coefficient: 1.0 or 0.5 Original image dimension: 512 x 512
- 5. Practical issue #1 True_Y image
- 6. Practical issue #2 Hounsfield unit
- 7. Practical issue #2 Hounsfield unit 데이터 비식별화 과정에서 실수가 일어남.
- 8. 2.1 Unet Application 2015 awards(they won the ISBI cell tracking challenge 2015 in these categories by a large margin) Segmentation golden standard Application Cells Segmentation Brain Segmentation MRBrains Challenge
- 9. 2.2 Information of in-house U-net(1) Convolution block ReLU activation function Padding Same Activation function at final layer Sigmoid function Cost function Dice loss function Cross Entropy function Optimizer Adam optimizer
- 10. 2.2 Information of in-house U-net(2) input size: 256x256 batch size: 10 learning rate: 0.0001 Learning time: 3 ~ 4hour 500epochs The number of Feature map: 2 4 8 16 32
- 11. 3.1 Good result: (1) onlyroi, dice, nii
- 12. 3.1 Bad result: (1) onlyroi, dice, nii
- 13. 3.1 Good result: (2) onlyroi, dice, jpg
- 14. 3.1 Bad result: (2) onlyroi, dice, jpg
- 15. 3.1 Good result: (3) onlyroi, sigmoid, jpg
- 16. 3.1 Bad result: (3) onlyroi, sigmoid, jpg
- 17. 3.1 Good result: (4) onlyroi, sigmoid, jpg
- 18. 3.1 Bad result: (4) onlyroi, sigmoid, jpg
- 19. Slice: only roi Loss: dice Dtype: nii Slice: only roi Loss: dice Dtype: jpg Slice: only roi Loss: sigmoid Dtype: nii Slice: only roi Loss: sigmoid Dtype: jpg
- 20. Cost graph
- 21. Practical issue #3 Accuracy metric 0.1 0.05 0.1 0.03 0.04 0.05 0.32 0.22 0.4 0.02 0.4 0.93 0.92 0.63 0.2 0.32 0.96 0.89 0.99 0.5 0.02 0.01 0.04 0.05 0.07 0 0 0 0 0 0 0 0 0 0 0 1 1 0 0 0 1 1 1 0 0 0 0 0 0
- 22. Practical issue #4 Prediction error In imbalance condition
- 23. Information among frames == Information among Slice image Practical issue #4 Prediction error In imbalance condition
- 28. 5. Proposed method Architecture
- 29. Article # segmentation Cai, Jinzheng, et al. "Improving deep pancreas segmentation in ct and mri images via recurrent neural contextual learning and direct loss function." arXiv preprint arXiv:1707.04912 (2017). Kim, Jung Uk, Hak Gu Kim, and Yong Man Ro. "Iterative deep convolutional encoder-decoder network for medical image segmentation." Engineering in Medicine and Biology Society (EMBC), 2017 39th Annual International Conference of the IEEE. IEEE, 2017. Oktay, Ozan, et al. "Anatomically Constrained Neural Networks (ACNN): Application to Cardiac Image Enhancement and Segmentation." arXiv preprint arXiv:1705.08302 (2017). 29
- 30. Papers Authors Journal Architecture #1 Improving deep pancreas segmentation in ct and mri images via recurrent neural contextual learning and direct loss function." Cai, Jinzheng, et al. arXiv / 2017 #2 "Iterative deep convolutional encoder- decoder network for medical image segmentation." Kim, Jung Uk et al IEEE / 2017 #3 "Anatomically Constrained Neural Networks (ACNN): Application to Cardiac Image Enhancement and Segmentation." Oktay, Ozan, et al. arXiv / 2017
- 31. Papers Authors Journal Architecture #1 Improving deep pancreas segmentation in ct and mri images via recurrent neural contextual learning and direct loss function." Cai, Jinzheng, et al. arXiv / 2017 #2 "Iterative deep convolutional encoder- decoder network for medical image segmentation." Kim, Jung Uk et al IEEE / 2017 #3 "Anatomically Constrained Neural Networks (ACNN): Application to Cardiac Image Enhancement and Segmentation." Oktay, Ozan, et al. arXiv / 2017
- 32. Papers Authors Journal Architecture #1 Improving deep pancreas segmentation in ct and mri images via recurrent neural contextual learning and direct loss function." Cai, Jinzheng, et al. arXiv / 2017 #2 "Iterative deep convolutional encoder- decoder network for medical image segmentation." Kim, Jung Uk et al IEEE / 2017 #3 "Anatomically Constrained Neural Networks (ACNN): Application to Cardiac Image Enhancement and Segmentation." Oktay, Ozan, et al. arXiv / 2017
- 33. Papers Authors Journal Architecture #1 "Improving deep pancreas segmentation in ct and mri images via recurrent neural contextual learning and direct loss function." Cai, Jinzheng, et al. arXiv / 2017 #2 "Iterative deep convolutional encoder- decoder network for medical image segmentation." Kim, Jung Uk et al IEEE / 2017 #3 "Anatomically Constrained Neural Networks (ACNN): Application to Cardiac Image Enhancement and Segmentation." Oktay, Ozan, et al. arXiv / 2017
- 34. Papers Authors Journal Architecture #1 "Improving deep pancreas segmentation in ct and mri images via recurrent neural contextual learning and direct loss function." Cai, Jinzheng, et al. arXiv / 2017 #2 "Iterative deep convolutional encoder- decoder network for medical image segmentation." Kim, Jung Uk et al IEEE / 2017 #3 "Anatomically Constrained Neural Networks (ACNN): Application to Cardiac Image Enhancement and Segmentation." Oktay, Ozan, et al. arXiv / 2017
- 35. Papers Authors Journal Architecture #1 "Improving deep pancreas segmentation in ct and mri images via recurrent neural contextual learning and direct loss function." Cai, Jinzheng, et al. arXiv / 2017 #2 "Iterative deep convolutional encoder- decoder network for medical image segmentation." Kim, Jung Uk et al IEEE / 2017 #3 "Anatomically Constrained Neural Networks (ACNN): Application to Cardiac Image Enhancement and Segmentation." Oktay, Ozan, et al. arXiv / 2017
- 36. Papers Authors Journal Architecture #1 "Improving deep pancreas segmentation in ct and mri images via recurrent neural contextual learning and direct loss function." Cai, Jinzheng, et al. arXiv / 2017 #2 "Iterative deep convolutional encoder- decoder network for medical image segmentation." Kim, Jung Uk et al IEEE / 2017 #3 "Anatomically Constrained Neural Networks (ACNN): Application to Cardiac Image Enhancement and Segmentation." Oktay, Ozan, et al. arXiv / 2017
- 37. Papers Authors Journal Architecture #1 "Improving deep pancreas segmentation in ct and mri images via recurrent neural contextual learning and direct loss function." Cai, Jinzheng, et al. arXiv / 2017 #2 "Iterative deep convolutional encoder- decoder network for medical image segmentation." Kim, Jung Uk et al IEEE / 2017 #3 "Anatomically Constrained Neural Networks (ACNN): Application to Cardiac Image Enhancement and Segmentation." Oktay, Ozan, et al. arXiv / 2017
- 38. Department of health Science and Technology, SAIHST, Sungkyunkwan University, Seoul 06351, Korea Liver Segmentation Liver segmentation using U-net; Practical issues
- 41. Liver segmentation using U-net