Fundus Disease Image Classification based on Improved Transformer

In recent years, the progress of deep learning and fundus camera technology makes it possible to diagnose fundus diseases by computer. However, the fundus image dataset is relatively small, which makes the pure transformer model challenging to be applied to medical disease analysis. Therefore, this paper proposes a Transformer Eye (TransEye) fine-grained fundus disease image classification method based on the self-attention mechanism to assist diagnosis. TransEye combines the advantages of Convolution Neural Network (CNN) and Transformer model. It can not only effectively extract the underlying features, but also establish the remote dependence of the image. So, it can locate the most discriminative image area and complete end-to-end training. Evaluated the classification effect of our method on the preprocessed OIA dataset, the experimental results show the superiority of TransEye.