SP-DCNet: A Support-Set Guided Directional Connectivity Network for Few-Shot Medical Image Segmentation

Few-shot medical image segmentation is crucial for diagnosis, treatment planning, and research, yet deep learning models struggle with limited annotated data and generalization to unseen categories. To address this, metalearning-based few-shot learning has emerged, enabling segmentation with minimal labeled samples. This paper proposes SP-DCNet, a novel approach that enhances segmentation accuracy by refining global features using low-dimensional directional information from the support set. The key contributions are: (1) the SPDE module, which utilizes prototype directional information for feature calibration; (2) the RPL block and SDE block, which integrates directional and global features; and (3) experiments on the CHAOS dataset, where SP-DCNet surpasses existing methods. This work demonstrates the effectiveness of combining meta-learning with low-dimensional information for few-shot medical image segmentation.