Abstract:Automatic detection of diabetic retinopathy (DR) by fundus images can help a lot of patients to get diagnosis. However, many excellent deep learning models for image classification do not equally achieve good results on DR detection since fundus image classification is different from typical image classification problem. The lesions are usually small and reflected in the part rather than the whole image. In this paper, we investigate the effect of network depth and the relationship between different DR grades. Then we propose a novel attention network to group and classify DR. Our contributions are mainly in three aspects: 1. Find a suitable network depth for DR detection by experiments. 2.Propose a grouping and decoupling mechanism for DR grading. 3. Design a novel deep convolutional neural network to classify DR, which also generates attention map indicating the location of lesions. Experiments show that our network outperforms other methods on the Messidor dataset. We reach the AUC of 96.9% and the accuracy of 92.7% on binary classification and the accuracy of 81.7% on 4-class grading.

Grouping and Decoupling Mechanism for Diabetic Retinopathy Image Grading.