BACKGROUND: Early and reliable grading of diabetic retinopathy is important for preventing avoidable vision loss. Although deep learning methods have shown strong performance on retinal fundus images, grading remains difficult because of class imbalance, variable image quality, subtle differences between adjacent disease stages, and the need for interpretable predictions. This study developed a clinically interpretable deep learning pipeline for five-class diabetic retinopathy grading. METHODS: A classification framework based on EfficientNet-B1 was developed using contrast-limited adaptive histogram equalization for image enhancement, sample-mixing augmentation, focal loss, and test-time augmentation. Model behavior was further examined using calibration analysis, gradient-weighted class activation mapping, and t-distributed stochastic neighbor embedding. The public APTOS 2019 dataset was used. Under the fixed held-out protocol, 2,930 images were used for training, 364 for validation, and 366 for testing. Robustness was further examined using stratified 5-fold cross-validation on a 3,294-image development set with an independent 366-image test set. Performance was summarized using accuracy, macro F1-score, weighted F1-score, average area under the receiver operating characteristic curve, confusion matrices, and calibration error. No formal hypothesis-testing framework was applied. RESULTS: The best validation accuracy under the fixed split was 84.07%. On the independent test set, the proposed model achieved 83.61% accuracy without test-time augmentation and 84.97% with test-time augmentation. The average area under the receiver operating characteristic curve was 90.40%. Removing contrast-limited adaptive histogram equalization reduced test accuracy to 80.87%, supporting its contribution to performance. The model showed improved robustness to class imbalance and provided clinically meaningful visual explanations, although Severe diabetic retinopathy remained the most challenging category. CONCLUSIONS: The proposed pipeline achieved competitive benchmark performance for five-class diabetic retinopathy grading and combined classification accuracy with supportive calibration and interpretability analyses. The method appears suitable as a research-stage decision-support approach, but the remaining difficulty in advanced disease grading and the use of a single public dataset indicate that external validation is still required before clinical deployment. TRIAL REGISTRATION: This study used a publicly available benchmark dataset and did not involve a randomized controlled trial or prospective enrollment of human participants.
Ghadai et al. (Tue,) studied this question.