CORE: Suppressing Spurious Similarity via Confidence-Aware Prototypes for Few-Shot Medical Image Segmentation

Few-shot medical image segmentation (FSMIS) aims to segment unseen anatomical structures using only a few annotated examples, alleviating the heavy annotation burden in clinical practice. Most existing FSMIS methods adopt prototype-based learning but suffer from two limitations. First, equal treatment of foreground pixels ignores their heterogeneous reliability and weakens prototype discriminability. Second, query mask prediction uses simple prototype-query similarity matching, which can induce spurious high similarity and mistakenly match background regions to the foreground prototype, producing false positives. In this work, we propose CORE, a confidence-aware prototype learning framework designed to suppress spurious similarity for FSMIS to address these issues. CORE constructs foreground prototypes from regions with different confidence levels, which prevents excessive averaging and preserves intra-class diversity under extremely limited supervision. Furthermore, an adaptive veto guided query prototype interaction module suppresses unreliable prototype-query matching under spurious inter-class similarity. Extensive experiments demonstrate that CORE achieves the best average performance across four few-shot medical segmentation datasets, with pronounced gains on CHAOS-MRI, Synapse-CT, and CMR, and consistent gains on Prostate-MRI. Our code is available at https://gitlab.com/xuchuanzhen/core.