a probabilistic u net for segmentation of ambiguous images
play

A Probabilistic U-Net for Segmentation of Ambiguous Images Simon A. - PowerPoint PPT Presentation

A Probabilistic U-Net for Segmentation of Ambiguous Images Simon A. A. Kohl 1*,2 , Bernardino Romera-Paredes 1 , Clemens Meyer 1 , Jeffrey De Fauw 1 , Joseph R. Ledsam 1 , Klaus H. Maier-Hein 2 , S. M. Ali Eslami 1 , Danilo Jimenez Rezende 1 , Olaf


  1. A Probabilistic U-Net for Segmentation of Ambiguous Images Simon A. A. Kohl 1*,2 , Bernardino Romera-Paredes 1 , Clemens Meyer 1 , Jeffrey De Fauw 1 , Joseph R. Ledsam 1 , Klaus H. Maier-Hein 2 , S. M. Ali Eslami 1 , Danilo Jimenez Rezende 1 , Olaf Ronneberger 1 1 DeepMind 2 German Cancer Research Center *work done during an internship at DeepMind

  2. A Probabilistic U-Net for Segmentation of Ambiguous Images Simon A. A. Kohl 1*,2 , Bernardino Romera-Paredes 1 , Clemens Meyer 1 , Jeffrey De Fauw 1 , Joseph R. Ledsam 1 , Klaus H. Maier-Hein 2 , S. M. Ali Eslami 1 , Danilo Jimenez Rezende 1 , Olaf Ronneberger 1 1 DeepMind 2 German Cancer Research Center *work done during an internship at DeepMind Poster #127 Medical Imaging Workshop Talk: Sat, Dec 8, 9:45 am

  3. Images are often Ambiguous 3

  4. Images are often Ambiguous Potential Expert Graders Cancer 4

  5. Images are often Ambiguous Potential Expert Graders Segmentations from our model (U-Net + conditional VAE) Cancer 5

  6. Deterministic U-Net Inference Image U-Net 6

  7. Probabilistic U-Net Sampling ๐›Ž , ๐žƒ prior Prior Net Latent Space Image U-Net 7

  8. Probabilistic U-Net Sampling ๐›Ž , ๐žƒ prior 1 Sample 1 * Prior Net Latent Space Image U-Net 8

  9. Probabilistic U-Net Sampling ๐›Ž , ๐žƒ prior 2 1 Sample 1 * Prior Net Latent Space 2 Image U-Net 9

  10. Probabilistic U-Net Sampling ๐›Ž , ๐žƒ prior 2 1 Sample 3 1 * Prior Net Latent Space 2 Image 3 U-Net 10

  11. Probabilistic U-Net Training ๐›Ž , ๐žƒ prior Prior Net Sample ๐ด Cross- Entropy Image Sample Groundtruth U-Net 11

  12. Probabilistic U-Net Training Position in Latent Space for this GT example? ๐›Ž , ๐žƒ prior Prior Net Sample ๐ด Cross- Entropy Image Sample Groundtruth U-Net 12

  13. Probabilistic U-Net Training Posterior Net ๐›Ž , ๐žƒ post ๐›Ž , ๐žƒ prior KL Prior Net Latent Space Sample ๐ด Cross- Entropy Image Sample Groundtruth U-Net 13

  14. Latent Space Analysis Probabilistic U-Net Image 14

  15. Latent Space Analysis Probabilistic U-Net Image 15

  16. Latent Space Analysis Probabilistic U-Net Graders 0 Image 1 2 3 16

  17. Lung Abnormalities Segmentation: Quantitative Results Energy distance ( lower is better ) 17

  18. Lung Abnormalities Segmentation: Quantitative Results Energy distance ( lower is better ) 1 18

  19. Lung Abnormalities Segmentation: Quantitative Results Energy distance ( lower is better ) 1 4 19

  20. Lung Abnormalities Segmentation: Quantitative Results Energy distance ( lower is better ) 1 4 8 20

  21. Lung Abnormalities Segmentation: Quantitative Results Energy distance ( lower is better ) 1 4 8 16 21

  22. Lung Abnormalities Segmentation: Quantitative Results Energy distance ( lower is better ) 22

  23. Lung Abnormalities Segmentation: Quantitative Results Energy distance ( lower is better ) 23

  24. Lung Abnormalities Segmentation: Quantitative Results Energy distance ( lower is better ) 24

  25. Cityscapes segmentation: Qualitative Results Input Image Ground-truth Grader Samples (Probabilistic U-Net) Styles stochastic flips: sidewalk sidewalk 2 47 % person 41 % person 2 35 % car car 2 veget. 29 % veget. 2 road road 2 24 % 25

  26. Cityscapes segmentation: Qualitative Results Input Image Ground-truth Grader Samples (Probabilistic U-Net) Styles stochastic flips: sidewalk sidewalk 2 47 % person 41 % person 2 35 % car car 2 veget. 29 % veget. 2 road road 2 24 % 26

  27. Cityscapes segmentation: Quantitative Results 27

  28. Conclusions โ— Learn conditional probability over segmentation maps โ— Each sample is a valid & consistent segmentation โ— The likelihoods are well calibrated โ— Works on large-scale, real-world data โ— Can also be trained with a uni-modal GT โ— Can be used to asses annotations under the model code: github.com/SimonKohl/probabilistic_unet 28

  29. A Probabilistic U-Net for Segmentation of Ambiguous Images Simon A. A. Kohl 1*,2 , Bernardino Romera-Paredes 1 , Clemens Meyer 1 , Jeffrey De Fauw 1 , Joseph R. Ledsam 1 , Klaus H. Maier-Hein 2 , S. M. Ali Eslami 1 , Danilo Jimenez Rezende 1 , Olaf Ronneberger 1 1 DeepMind 2 German Cancer Research Center *work done during an internship at DeepMind Poster #127 Medical Imaging Workshop Talk: Sat, Dec 8, 9:45 am

  30. Probabilistic Segmentation: Clinical Use-Cases โ— Best-fit could be picked by clinician and adjusted if necessary. โ— Hypotheses could be propagated into next diagnostic pipeline steps. โ— Hypotheses could inform actions to resolve ambiguities. 30

  31. Evaluation Metric for Quantitative Comparison We use the Energy Distance 1 statistic (aka MMD): P gt P out where d(x,y) = 1 - IoU(x,y) and 1 Szรฉkely, G.J., Rizzo, M.L.: Energy statistics: A class of statistics based on distances. Journal of statistical planning and inference 143(8) (2013) 1249โ€“1272 31

  32. Baselines Normal Prior 1 1 2 1 3 2 2 Sample ๐ด 1 , ๐ด 2 , ๐ด 3 ,... 1,2,3,... m m U-Net Dropout U-Net M-Heads Image2Image VAE U-Net Ensemble 32

Recommend


More recommend