Correction of medical image segmentation errors through contrast learning with multi-branch

Tianlei Gao, Lei Lyu, Nuo Wei, Tongze Liu, Minglei Shu, Yushui Geng

Abstract: Medical image segmentation plays a pivotal role in computer-aided diagnosis. Despite the considerable advancements achieved by convolutional neural networks, segmentation errors such as under-segmentation and mis-segmentation remain significant obstacles, particularly in complex medical images where target regions often share similar features with the background. In this paper, we propose a generalized network called Correction of Medical Image Segmentation Errors through Contrast Learning with Multi-Branch Network (MEC) to address these biases and ultimately improve segmentation accuracy. MEC leverages a multi-branch architecture, where each branch is specifically designed to tackle distinct segmentation issues. The negative branch focuses on encoding mis-segmented regions, the positive branch captures accurately labeled regions, and the anchor branch refines the segmentation output from the base network. To effectively integrate the complementary information from these branches, we employ a contrastive learning strategy that aligns the semantic features of the anchor branch with those of the positive branch, while maintaining a clear distinction from the negative branch. This methodology enhances segmentation performance by capitalizing on the unique strengths of each branch. Experimental results on several medical image datasets demonstrate significant improvements in segmentation accuracy, thereby validating the effectiveness of our proposed approach.