A Two-Stage Deep Learning Cascade for Multi-Label Retinal Detection and DR Grading
Abstract
Diabetic Retinopathy (DR) is a leading cause of blindness among diabetic individuals, where early detection is critical yet hindered by a global shortage of ophthalmologists, particularly in resource-constrained regions. This paper proposes a two-stage cascade pipeline for automated multi-label retinal disease detection and DR severity grading. Stage 1 employs EfficientNetV2-S with Asymmetric Loss (ASL) on the RFMiD dataset to classify 28 retinal disease categories, achieving a mean AUC of 0.910 and mAP of 0.539. DR-positive cases are routed via a Platt-calibrated uncertainty-aware gate to Stage 2, which applies ordinal MSE regression on APTOS-2019, yielding a fivefold ensemble Quadratic Weighted Kappa (QWK) of 0.9339. Two novel contributions include a co-occurrence degradation analysis - revealing that multi-disease images suffer significant performance drops compared to single-disease cases - and a cascade gate that achieves QWK of 0.904 while preserving full 28-disease detection output. The modular pipeline demonstrates robust cross-domain transfer and is designed for deployment in low-resource clinical screening settings.