Leveraging spatial prior and grid attention for few-shot medical segmentation

In trained few-shot segmentation, an inevitable bias towards the base classes may occur. By leveraging the spatial consistency of organ distribution in medical image segmentation, spatial priors can be utilized to roughly predict the location of new categories. This helps to avoid bias towards the base classes. This paper introduces a novel prototype network that integrates prior spatial knowledge and grid attention to enhance the accuracy of few-shot medical image segmentation by predicting target locations and offering guidance for segmentation algorithms. We introduce dedicated spatial prior module to extract information about the location of targets. This module utilizes the spatial and shape consistency of supported and query targets in medical images to better guide the segmentation process. By incorporating a spatial prior module, the segmentation network is guided to follow the desired spatial structure of the target. Additionally, to emphasize essential local semantic information in medical images, we have devised a Grid Attention (GA) module. The GA module adaptively adjusts the importance of different positions in the feature map, allowing more attention to be placed on regions relevant to the segmentation task. By combining spatial priors with grid attention, our method can better capture target information in medical images and achieve more accurate segmentation results. This claim was substantiated through our experiments focusing on segmenting abdominal organs in MRI scans. Our method was objectively compared against several established models, demonstrating that the inclusion of spatial priors and grid attention greatly improved the segmentation results.