Boundary-Aware Few-Shot Framework Improves Medical Image Segmentation Accuracy

Across extensive experiments conducted on three publicly available medical image datasets, the decoupled model demonstrated consistent improvements over previous approaches, achieving mean Dice similarity scores of 84.37 percent on CHAOS-MRI, 80.09 percent on Synapse-CT, and 79.85 percent on MS-CMRSeg.

Few-shot segmentation models have gained great popularity in the medical image segmentation community. Most existing methods often encode support and query images to feature maps, generating prototypes from support features and labels, and comparing query features to these prototypes to identify regions of interest. Under these conventional pipelines, the low-resolution segmentation result generated from feature maps via prototype matching is directly upsampled.

To address this problem, the researchers decoupled the few-shot segmentation pipeline into two complementary components: a prototypical localization module and an adaptive boundary refinement module. The prototypical localization module extracts semantic prototypes to locate target tissues within the query images.

Additionally, a novel self-supervised learning task is introduced to train those two modules effectively. Instead of learning only one single sampled pseudo mask, the self-supervised learning strategy learns all valid pseudo-mask pairs generated by supervoxel partitioning. The authors noted that this strategy allows the model to better capture global semantic information, making the model learn the image data more effectively and comprehensively than approaches that rely on a single sampled pseudo mask.

Subject of Research: Boundary-aware few-shot medical image segmentation

Article Title: Boundary-aware few-shot medical image segmentation with refinement module and self-supervised learning

Article References: Tu, S., & Chai, L. (2026). Boundary-aware few-shot medical image segmentation with refinement module and self-supervised learning. Applied Intelligence, 56(16), Article 551. https://doi.org/10.1007/s10489-026-07590-7

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07590-7

Keywords: medical image segmentation, few-shot learning, boundary refinement, self-supervised learning, supervoxel partitioning, CHAOS-MRI, Synapse-CT, Boundary-aware, few-shot, medical, image, segmentation