MOrphologically-Aware Jaccard-Based ITerative Optimization (MOJITO) for Consensus Segmentation

The extraction of consensus segmentations from several binary or probabilistic masks is important to solve various tasks such as the analysis inter-rater variability fusion neural network outputs. One most widely used method obtain a segmentation STAPLE algorithm. In this paper, we first demonstrate that output algorithm heavily impacted by background size images and choice prior. We then propose new construct based on Fréchet means Jaccard distances which make it totally independent image size. provide heuristic approach optimize criteria voxel’s class fully determined its morphological distance, connected component belongs group raters who segmented it. compared extensively our three datasets with naive averaging method, showing leads intermediate between Majority Voting different posterior probabilities than those methods. Codes are available at https://gitlab.inria.fr/dhamzaou/jaccardmap .