LiteDR-XAI: A Lightweight and Explainable Framework for Referable Diabetic Retinopathy Screening

Automated diabetic retinopathy screening must balance predictive performance with computational cost, threshold control, reproducibility, and interpretability. This paper presents LiteDR-XAI as a reproducible framework for binary referable diabetic retinopathy screening on APTOS 2019, not as a new con volutional architecture or a universal state-of-the-art model. Five pretrained backbones are evaluated through 35 controlled runs covering input resolution, contrast-limited adaptive histogram equalization, pretraining, backbone freezing, and loss design. Checkpoints are selected by validation AUROC, and the operating threshold is estimated only from validation predictions using Youden’s index before application to the held-out test set. From these experiments, a compact MobileNetV3-Large configuration at 320 pixels is selected as the LiteDR-XAI operating model because it provides a balanced accuracy-efficiency-explainability trade-off: 94.73% accuracy, 0.9791 AUROC, 0.9646 AUPRC, 95.96% sensitivity, 93.88% specificity, 91.45% precision, and 93.65% F1-score with 3.22 million parameters. A strict literature comparison includes only near-direct APTOS binary/referable DR studies and shows that LiteDR-XAI has the highest accuracy among the included comparisons, while other methods report higher AUC or sensitivity under different protocols. Grad CAM supports qualitative inspection of correct and erroneous predictions. External, multi-seed, device-level, and prospective validation remain necessary before deployment.