Mixed Informed Transformer for Few-Shot Medical Image Segmentation

With the rapid development of deep learning techniques trained on datasets that require a large amount of manual annotation, significant progress has been made in the field of medical image segmentation. However, the annotation of medical image analysis requires not only very high labor costs, but also specific expertise in the medical field. In order to address this limitation, few-shot medical image segmentation has become a very promising and popular research direction, aiming to learn novel classes from limited supervised medical data. We propose a novel framework called Mixed Informed Transformer (MIT) for few-shot medical image segmentation. Our MIT feasibly augments the representation capability of support prototype and query features. By incorporating augmented query features and support features, MIT effectively learns the foreground information of medical images and eliminates redundant background information. Experimental results obtained on three benchmark datasets are compared with state-of-the-art few-shot medical image segmentation methods, which demonstrates the effectiveness and competitiveness of our proposed approach.