Diabetic Retinopathy Detection Method Based on Improved Convolutional Neural Network Using Fine-Tuning

The purpose of the paper is to apply the improved convolutional neural network to help detect diabetic retinopathy, which requires us to classify the degree of diabetic lesions into five categories through the pictures of the fundus of the eyes: Normal, Mild, Moderate, Severe, Proliferative. Based on the tensorflow framework, we built a Resnet34 model, and used image enhancement and fine-tuning to improve the effect of the model, and finally achieved an accuracy of about 83% on the test set. We also compared the effects of other network structures. The accuracy of the Densenet121 model on the test set is 82%, Inception V3 is 81%, and Alexnet is 71%.