Self-supervised exceptional prototypical network for few-shot grading of gastric intestinal metaplasia

Automatic grading of Gastric Intestinal Metaplasia (GIM) is valuable in assisting the diagnosis of early gastric cancer. Recently, prototypical networks are served as a effective method for medical image processing in few-shot scenarios. However, existing prototypical networks suffer from the following two limitations when applied to GIM grading: 1) Variable camera angles of gastric endoscopes result in diverse sampling granularities of GIM lesions, leading to a multitude of multiscale features. Fully supervised encoders struggle to learn robust multiscale features due to limited labeled endoscopic images and privacy concerns. 2) Class prototypes based on sample means ignore the latent class information of exceptional cases, resulting in one-sided inferences of category prototypes and decision boundaries. To address these challenges, we propose a Self-supervised Exceptional Prototypical Network (Swin-EPN) for few-shot grading of GIM. Specifically, three tailored pretext tasks are designed to jointly pretrain a swin transformer, which is integrated as the model's embedding layer to learning robust multiscale features. We propose an exceptional prototype mining module that identifies exceptional prototypes by defining a prototype score for each sample and updating potential exceptional prototypes in an exceptional prototype bank. These exceptional prototypes are served as supplementary information to class prototypes, and are leveraged to guide the delineation of class decision boundaries. We validated Swin-EPN on a private GIM dataset from a local grade-A tertiary hospital in both 1-shot and 5-shot scenarios, achieving accuracy improvements of 6.12% and 5.61% respectively compared to state-of-the-art (SOTA) models.