Self-Guided Local Prototype Network for Few-Shot Medical Image Segmentation

Recently, few-shot medical image segmentation approaches have been extensively explored to tackle the challenge of scarce labeled data in medical images. The majority of existing methods employ prototype-based techniques and have achieved promising results. However, conventional prototype extraction approaches inherently lead to loss of spatial information, thus degrading model performance, an issue further aggravated in medical images with large background regions. In this work, we propose a self-guided local prototype generation module (SLP), which progressively splits support masks into sub-mask, thereby producing a set of local prototype that preserve richer support image information. Moreover, in order to take full advantage of the information contained within the prototype sets during the iterative process, we generate a prior mask from this information and provide coarse spatial location about the target for the model through a simple prior-guided attention module (PGA). Experiments on three different datasets validate that our proposed approach outperforms existing methods.