Advancing Few-Shot Medical Image Segmentation: Multi-Prototype Guidance via Signed Distance
Yunfei Gao, Haifeng Zhao · 2024
Few-shot medical image segmentation presents a class-agnostic training solution to address the critical challenge of scarce data annotations in medical image data. Nevertheless, most existing methods based on prototype learning use windowing to obtain multiple sub-regions to extract prototypes, resulting in the foreground and background region features being entangled and not being efficiently utilized. In this paper, we propose a novel local prototype method guided by signed distance to tackle the challenge of feature entanglement. First, we employ signed distance guidance for extracting multiple prototypes to ensure absolute utilization of all foreground features. Subsequently, similarity comparison yields initial predictions. Finally, we use the shape loss to help the model achieve better results in shape prediction. Extensive evaluations conducted on two medical image benchmark datasets, including CHAOs and Synapse, demonstrate that our model achieves outperformance compared to existing state-of-the-art methods.