Diabetic Retinopathy Grading Using Attention based Convolution Neural Network
Zhixiang Qian, Chenjian Wu, Hong Chen, Minxin Chen
Abstract
In this paper, an automatic diagnosis method based on deep learning algorithm is proposed, which will speed up the diagnosis of diabetic retinopathy and improve the efficiency of treatment. We built a convolutional neural network model called as "AD2Net". The network combines the advantages of Res2Net and DenseNet, which can not only learn multi-scale features, but also alleviate the vanishing-gradient problem and strengthen feature reuse. At the same time, this paper also uses attention mechanism method to encourage the network to focus on learning useful information in images, which can improve classification effect of the network to a certain extent. The results show that the method proposed in this paper can divide the fundus images into five stages of disease based on severity. The accuracy and Kappa values that the model has achieved are 83.2% and 0.8 respectively on testing set. Compared with the existing method, the method proposed in this paper has certain advantages.