Beyond singular prototype: A prototype splitting strategy for few-shot medical image segmentation

In the realm of medical image semantic segmentation, few-shot learning, characterized by its efficient data utilization and flexible generalization capabilities, has been garnering increasing attention. The mainstream methods currently employ prototype-based approaches, which extract semantic knowledge from the annotated support images to guide the segmentation of the query image via masked global average pooling. However, such masked global average pooling leads to severe information loss, which is more problematic for medical images with large numbers of highly heterogeneous background categories. In this work, we propose a prototype splitting module (PSM) to effectively address the issue of semantic information loss in few-shot medical image segmentation. Specifically, PSM iteratively splits the support image masks into set of sub-masks containing segmented regions and unsegmented regions in a self-guided manner. This maximally retains the information within the original semantic classes and better extracts the representations of those classes. Additionally, we devise a multi-level cross attention module (MCAM) that transfers the foreground information from the support images to the query images across different levels to facilitate final segmentation prediction. We validate our method on multiple modal and multi-semantic medical image datasets. Results demonstrate that our approach achieves superior performance over existing state-of-the-art methods. The code has been released on https://github.com/fdngh/PSMnet.