Diabetic retinopathy detection: custom CNN architecture with regularization and data augmentation for improved generalization and efficiency

Abstract

Diabetic Retinopathy (DR) is still among the leading causes of blindness in the working age population worldwide. Manual screening is inconvenient and not feasible. Current research proposes deployment of a light CNN, which was trained with the ODIR dataset (5,018 fundus images: 2,574 Normal and 2,444 Abnormal/DR), on a 70:15:15 train-validation-test split. Dynamic class weights implemented within Binary Focal Loss handled imbalance- induced bias. Compared to a DenseNet121 baseline of 83% accuracy implemented with Test Time Augmentation, the proposed model achieved 98% accuracy and 0.98 (Normal) and 0.97 (Abnormal) F1-scores. Robustness was achieved with LAB color preprocessing and CLAHE enhancement, real-time data augmentation, and stratified sampling. The model’s efficiency enables Edge-AI deployment in low-resource environments. Future work will incorporate Explainable AI and multi-source validation to enhance interpretability and clinical reliability.