Few–shot medical image segmentation via prototype mining and prior–guided matching network

Few-shot medical image segmentation (FSMIS) aims to achieve accurate segmentation of target regions using only a very limited number of annotated samples. However, target organs in medical images often show substantial scale variations and pronounced intra-class heterogeneity, making it difficult for existing methods to learn stable and discriminative class representations. To address the above issues, we propose a Prototype Mining and Prior-guided Matching Network (PMPNet). Specifically, we introduce an Adaptive Prototype Mining (APM) module, which jointly models global prototypes and an adaptive number of local prototypes to construct multi-granularity hierarchical support representations, thereby more comprehensively describing the internal heterogeneity of foreground and background regions. In addition, we observe that the discriminative information from the support branch provides insufficient guidance for query feature learning, making query representations susceptible to complex background interference. To alleviate this issue, we develop a Prior-Driven Query Enhancement (PDQE) module, which explicitly incorporates prior information generated from the support set into the query feature learning process and combines it with a multi-scale hierarchical refinement mechanism to enhance target-region responses while suppressing irrelevant background interference. Furthermore, considering that different prototypes vary in representation quality and category representativeness, we devise a Multi-Prototype Matching Prediction (MPMP) module to improve foreground-background discrimination. Extensive experiments on the Abd-MRI, Abd-CT, and CMR medical image datasets demonstrate that the proposed method achieves competitive few-shot segmentation performance, and detailed analyses further verify its effectiveness.