An Improved Prototypical Network for Endoscopic Grading of Intestinal Metaplasia

Intestinal metaplasia (IM) is confirmed to be the commonest symptom of early gastric cancer. IM grading by endoscopic images is essential to reduce the risk and mortality of gastric cancer. However, expensive and time-consuming manual annotations result in insufficient endoscopic image datasets to support the training of deep learning models. Prototypical network is the foundational model in few-shot learning and routinely is used to overcome insufficient datasets. Therefore, we propose an improved prototypical network (PN-ViT) for endoscopic grading of intestinal metaplasia. Firstly, the original IM dataset is divided into support sets and query sets, where the support set images are used to calculate prototypes and the images are only used to extract features in the query set. Second, a pre-trained vision transformer (ViT) is used as an embedding layer to extract IM lesion features from support sets. Then, the average value of features is calculated as the prototype of each category. Third, the features of the query set images are extracted using the pre-trained ViT as well. After the distances between them and prototypes are calculated for predicting the category. Finally, we conducted extensive experiments on a private intestinal metaplasia grading dataset from a Grade-A tertiary hospital. Experimental results showed that PN-ViT achieved the top classification accuracy of 70.2% in 3-way 5-shot scenario and verified the effectiveness of PN-ViT.