Weakly-Supervised 3D Medical Image Segmentation using Geometric Prior and Contrastive Similarity

• DOI: 10.48550/arxiv.2302.02125
• Publication Date: 2023-01-01
• Authors: Hao Du; Qihua Dong; Yan Xu; Jing Liao
• Keywords: Robustness; Minimum bounding box; Similarity (geometry); Bounding overwatch

Medical image segmentation is almost the most important pre-processing procedure in computer-aided diagnosis but also a very challenging task due to complex shapes of segments and various artifacts caused by medical imaging, (i.e., low-contrast tissues, non-homogenous textures). In this paper, we propose simple yet effective framework that incorporates geometric prior contrastive similarity into weakly-supervised loss-based fashion. The proposed built on point cloud provides meticulous geometry proposal, which serves as better supervision than inherent property bounding-box annotation height width). Furthermore, encourage organ pixels gather around embedding space, helps distinguish tissues. space can make up for poor representation conventionally-used gray space. Extensive experiments are conducted verify effectiveness robustness framework. superior state-of-the-art methods following publicly accessible datasets: LiTS 2017 Challenge, KiTS 2021 LPBA40. We dissect our method evaluate performance each component.