Fig. 1. Method flowchart

Fig. 2. Example of the preprocessing of fundus images

Fig. 3. The dilated convolution corresponding to different expansion rates

Fig. 4. Examples of retinal fundus images

Fig. 5. Workflow diagram of the proposed approach for classifying fundus images (ACC: accuracy; SE: sensitivity; SP: specificity; AUC: area under the receiver operating characteristic curve)

Fig. 6. Multi-class comparison results

Fig. 7. Accuracy and loss on the training datasets

Fig. 8. Intermediate results of the experiment

Fig. 9. Image examples of experimental result

Fig. 10. Result of experiment operation

Fig. 11. Classification of DR images by multiple convolutional neural networks