DAPNet: a decoupled adaptive prototyping network for few-shot medical image segmentation

Few-shot medical image segmentation (FSMIS) holds immense potential in medical scenarios where annotations are expensive and scarce. However, the majority of existing methods perform queries under the guidance of a foreground prototype, which causes them to struggle, as this single representation fails to effectively model both the target and its complex anatomical surroundings. To address this limitation, we propose the Decoupled Adaptive Prototype Network (DAPNet), explicitly disentangling the representation of foreground and background through a Dual-Path Dynamic Feature Enhancement and Feedback Framework. The Deformable Anatomical Prototype Sampling (DAPS) module, which innovatively combines coordinate-aware and deformable convolutions to achieve adaptive anatomical alignment, resulting in a robust set of background prototypes derived from support features; and the Dual Attention-Boosted Prototyping (DABP) module, which generates a highly discriminative foreground prototype by fusing semantic attention with a boundary-aware mechanism to effectively mitigate feature bias and enhance the structural integrity of the target. By jointly optimizing these distinct but complementary prototypes, DAPNet forms a powerful and comprehensive representation for classification. Extensive experiments on multiple public datasets demonstrate that our method consistently outperforms state-of-the-art approaches. Ablation studies and visualizations further validate the effectiveness and synergy of our proposed modules, establishing a new and robust paradigm for FSMIS.