AFA-RN: An Abnormal Feature Attention Relation Network for Multi-class Disease Classification in gastrointestinal endoscopic images

Gastrointestinal diseases pose a great threat to human health. Early detection and treatment will significantly reduce the fatality rate. At present, wireless capsule endoscopy (WCE) is widely used in the clinical examination of gastrointestinal diseases. Each examination will produce tens of thousands of WCE images, only a small part of which are images with diseases. It is very time-consuming and laborious for doctors to read one by one to pick out the images with diseases, so the development of computer classification algorithm is very valuable. Previous deep learning-based methods require a large amount of labeled data, but it is often difficult to obtain such well-labeled data, because experienced experts and doctors are required to make the labeling to ensure the reliability of data labeling. In addition, there are few clinical data samples of some types of disease images, which leads to the problem of imbalance of samples in the disease classification data set of WCE images. To solve this problem, we proposed an abnormal feature attention relationship network -- AFA-RN, which used feature addition, feature concatenation and bilinear merge to build the abnormal feature attention module, so that the few-shot learning relationship network model could greatly improve its performance in the multi-category disease classification task of gastrointestinal images.