Multi-stage classification of diabetic retinopathy based on mixed attention network

Diabetic retinopathy (DR) has small differences among the multi-stage types, it is difficult to classify accurately in early stage. For this reason, we proposed a high-efficiency model of diabetic retinopathy based on hybrid attention and multi-stage, which is named Coordinate Attention and Squeeze- and-Excitation Networks EfficientNet-V2 (CS-ENetV2). Firstly, we used coordinate attention mechanism and hybrid attention mechanism of compression and excitation network module to localize the feature target precisely. Secondly, we designed a hybrid attention efficient network based on diabetic retinopathy feature extraction to further optimize the quality of feature extraction. Finally, to enhance the accuracy of the classification, the feature information obtained from each layer is fused by using multi-feature fusion. Moreover, we used a migration learning strategy to pre-train the CS-ENetV2 model. After pre-processing the publicly available dataset from Kaggle, relevant classification experiments were done by us, the experimental results showed that the specificity, sensitivity, classification accuracy and Quadratic Weighted Kappa score of the model reached 97. 8%, 93.8%,94.9%and 0.86, which are better than the current typical classification methods and have better classification and robustness. Our method provides a good solution for accurate automatic staging of diabetic retinopathy.