Deployment Aligned Hybrid Attention For Diabetic Retinopathy Screening With Mdrnet

Early identification of diabetic retinopathy (DR) is critical for preventing avoidable vision loss, particularly in primary-care settings where screening infrastructure is constrained by variable image quality, limited hardware resources and workforce pressure. This work introduces MDRNet, a deployment-aligned deep learning framework designed to provide clinically coherent DR grading under these real-world conditions. The approach integrates F-CAN perceptual harmonisation, a lightweight EfficientNet-B0 encoder, multi-scale hybrid attention, and class-distribution regularisation, yielding a model tailored for commodity CPU hardware without reliance on transformer-scale capacity. Evaluation on a patient-level EyePACS split that preserves natural epidemiological imbalance demonstrates robust multiclass performance, achieving 91.3% accuracy, 0.86 macro-F1, 0.90 quadratic weighted kappa, and 0.95 macro-AUC across the five-grade clinical scale. Referral-tier mapping, which collapses severe and proliferative disease into a clinically actionable category, achieved 0.92 sensitivity and 0.85 specificity, indicating suitability as a triage-supportive system. Calibration metrics (Brier score = 0.082; ECE = 0.034) further suggest reliability of probabilistic outputs for threshold-based decision support. The findings demonstrate that prioritising calibration, ordinal stability and inference feasibility can yield competitive performance without specialised accelerators, supporting the potential for MDRNet to contribute meaningfully to scalable DR screening pathways subject to prospective validation.