CRISP - Reliable Uncertainty Estimation for Medical Image Segmentation

Accurate uncertainty estimation is a critical need for the medical imaging community. A variety of methods have been proposed, all direct extensions classification estimations techniques. The independent pixel-wise estimates, often based on probabilistic interpretation neural networks, do not take into account anatomical prior knowledge and consequently provide sub-optimal results to many segmentation tasks. For this reason, we propose CRISP ContRastive Image Segmentation Prediction method. At its core, implements contrastive method learn joint latent space which encodes distribution valid segmentations their corresponding images. We use compare predictions thousands vectors anatomically consistent maps. Comprehensive studies performed four image databases involving different modalities organs underlines superiority our compared state-of-the-art approaches. Code available at: https://github.com/ThierryJudge/CRISP-uncertainty .