A two-stage deep learning referral system with progressive training for diabetic retinopathy screening and severity grading.

Diabetic Retinopathy (DR) affects millions of people worldwide, but screening at population-scale is still limited by the lack of specialists and the clinical inefficiency of existing deep learning systems, which either perform binary referral or severity grading but not both within a single unified routed architecture. We present a shared EfficientNet-B4 backbone with two independently activatable heads: a binary screener (Head 1) and a joint five-class DR severity and three-class Diabetic Macular Edema (DME) risk grader (Head 2), trained by way of a five-stage progressive curriculum which prevents the gradients of one pathway from interfering with the other. Clinical routing only invokes the grader for referred cases, reducing computational demand as a function of DR prevalence. The screener achieves AUC =0.9789 on APTOS and AUC =0.8685 on zero-shot external validation (Messidor-2); a small calibration sample of 50 Messidor-2 images recovers ≥90% sensitivity, demonstrating that the AUC gap reflects threshold shift rather than a model failure. The grader achieves five-class DR severity QWK =0.7731 with simultaneous three-class DME risk output; an ordinal loss and MixUp remediation stage (Stage 5b) rebalances per-class performance, substantially improving No DR and Severe NPDR F1 at the cost of a partial reduction in PDR F1 and a decrease in Grade 1 F1 from 0.1818 to 0.1429. The principal limitation is that the grader was trained on only 413 IDRiD images, and Grade 1 F1 remains low across all variants, reflecting the scarcity of annotated grading data.