A Probabilistic U-Net for Segmentation of Ambiguous Images

  • 2018-06-13 13:47:04
  • Simon A. A. Kohl, Bernardino Romera-Paredes, Clemens Meyer, Jeffrey De Fauw, Joseph R. Ledsam, Klaus H. Maier-Hein, S. M. Ali Eslami, Danilo Jimenez Rezende, Olaf Ronneberger
  • 67

Abstract

Many real-world vision problems suffer from inherent ambiguities. In clinicalapplications for example, it might not be clear from a CT scan alone whichparticular region is cancer tissue. Therefore a group of graders typicallyproduces a set of diverse but plausible segmentations. We consider the task oflearning a distribution over segmentations given an input. To this end wepropose a generative segmentation model based on a combination of a U-Net witha conditional variational autoencoder that is capable of efficiently producingan unlimited number of plausible hypotheses. We show on a lung abnormalitiessegmentation task and on a Cityscapes segmentation task that our modelreproduces the possible segmentation variants as well as the frequencies withwhich they occur, doing so significantly better than published approaches.These models could have a high impact in real-world applications, such as beingused as clinical decision-making algorithms accounting for multiple plausiblesemantic segmentation hypotheses to provide possible diagnoses and recommendfurther actions to resolve the present ambiguities.

 

Quick Read (beta)

loading the full paper ...