<![CDATA[Diabetic retinopathy (DR) is a leading cause of preventable blindness among diabetes patients. This study optimizes DR severity classification using EfficientNet-B3 with transfer learning combined with data handling strategies. Using the APTOS 2019 dataset containing 3,662 retinal fundus images across five severity classes, three experimental scenarios were evaluated: (1) baseline CNN, (2) CNN with data augmentation, and (3) CNN with data augmentation and random oversampling. Performance was measured using Quadratic Weighted Kappa (QWK), accuracy, precision, recall, F1-score, and ROC-AUC. Results demonstrate that Scenario III achieves the best performance with QWK of 0.8496 and accuracy of 77.00%, representing significant improvement over baseline (QWK: 0.4998) and augmentation-only models (QWK: 0.5728). The combination of data augmentation and random oversampling effectively addresses class imbalance in medical image datasets. This study provides empirical evidence on combining transfer learning with data balancing strategies for automated DR screening systems.]]>
Copyrights © 2026
