Weakly-supervised 3d medical image segmentation using geometric prior and contrastive similarity

Abstract

A weakly-supervised framework for segmenting a three-dimensional (3D) medical image incorporates information of geometric prior and contrastive similarity in a loss-based fashion for enhancing distinguishability of an organ in presence of complex organ shape and image artifacts (e.g., low-contrast tissues) in the image. Training images with bounding-box annotations are used to train the framework. The geometric prior is advantageously built on point-cloud representations, which are much more fine-grained and flexible than volume representations, thus serving as better supervision than an inherent property of the bounding-box annotations (i.e. height and width). Furthermore, using contrastive similarity encourages organ pixels to gather around in a contrastive embedding space. It helps more-effectively distinguish low-contrast tissues from the background than using the conventionally-used gray space.