Classification of Diabetic Retinopathy Using ResNet50

Deep learning has been proposed as an automated solution for classifying the severity levels of Diabetic Retinopathy (DR). In this study, we utilized ResNet50 architecture to classify DR using the APTOS2019 dataset. As an initial step, we initialized the model with pre-trained weights from ResNet50 on ImageNet and implemented augmentation and resampling during training. We adopted an ensemble approach combined with classifiers such as SVM, Random Forest, and Logistic Regression, resulting in a ResNet50-Ensemble (SVM+RF+LR), with outputs obtained using a Soft Voting Classifier. The model achieved an accuracy of 85%, with a precision of 0.72, recall of 0.71, and F1-score of 0.71. The AUC values for the normal, mild, moderate, severe, and proliferative classes were 1.00, 0.96, 0.95, 0.95, and 0.91, respectively, with a Macro-average AUC of 0.96. These findings indicate that the appropriate use of ensemble methods can significantly enhance DR classification performance with suitable optimization strategies.