New AI pipeline grades diabetic retinopathy with 91.31% accuracy using EfficientNetV2-L

A new research paper from Saksham Kumar presents a unified deep learning pipeline for automated diabetic retinopathy (DR) grading that goes beyond raw accuracy to address real medical deployment challenges. The method combines Ben-Graham-green-channel CLAHE feature representation with an EfficientNetV2-L ordinal regressor and Monte Carlo dropout to provide calibrated uncertainty and referral decisions. Grad-CAM visual explanations are aligned with clinically relevant lesions, giving clinicians insight into model reasoning. On the APTOS-2019 official test split, the approach achieves a quadratic weighted kappa (QWK) of 91.31%, placing it in the near-perfect agreement range (>80%) when compared to human graders.

The system's practical safety mechanism sets it apart: at a 20% referral rate, 293 of 366 images are automatically graded with a QWK of 90.40%, while the remaining complex cases are escalated to specialists. This trade-off between automation and patient safety demonstrates a deployment-ready framework for clinical settings. The paper addresses the five-stage ordinal nature of DR severity, making the model more clinically aligned than simple classification. By combining uncertainty estimation, explainability, and ordinal regression, the work provides a template for trustworthy medical AI that can reduce specialist workload while maintaining reliability.