Generalized Probabilistic U-Net for Medical Image Segementation

We propose the Generalized Probabilistic U-Net, which extends U-Net [14] by allowing more general forms of Gaussian distribution as latent space that can better approximate uncertainty in reference segmentations. study effect choice has on capturing segmentations using LIDC-IDRI dataset. show affects sample diversity predictions and their overlap with respect to For dataset, we a mixture Gaussians results statistically significant improvement generalized energy distance (GED) metric standard U-Net. have made our implementation available at https://github.com/ishaanb92/GeneralizedProbabilisticUNet .