Self-attention light network for hierarchical severity detection of diabetic retinopathy

Abstract:
Diabetic retinopathy (DR) is a retinal disease caused by diabetes. It is one of the leading causes of permanent blindness. Therefore, clinical screening and classification of disease severity in patients diagnosed with diabetes are warranted. Currently, deep learning-based diabetic retinopathy screening tends to use deep networks such as Deep Convolutional Neural Network (DCNN). However, medical datasets usually contain fewer images. Thus, DCNN is not suitable for such small datasets. Light networks not only have fewer parameters that small datasets can suffice, but also they are more computationally efficient. They are more suitable for small datasets. Thus, in this paper, a novel light network for automatic DR detection is proposed. The proposed method is based on ResNet and attention modules. The attention modules are added to catch subtle features in images, and some convolutional layers are removed to prevent overfitting. Compared with existing DCNN-based DR detection methods, the proposed method has the following advantages. First, it has fewer parameters. Second, it optimizes the process layers. Third, the computing is convenience. Experimental results demonstrate that our proposed network outperforms other benchmark methods in accuracy, precision, recall reaching at 0.806, 0.814, 0.823 respectively and the time cost.