Detection of Diabetic Retinopathy using Convolutional Neural Network for Feature Extraction and Classification

Abstract

Sustained escalation of global diabetes incidence has rendered retinopathy-induced blindness a mounting public-health crisis, particularly among economically active adults whose occupational productivity depends on adequate vision. Although periodic fundus photography remains the cornerstone of population screening, its operational dependence on trained ophthalmologists limits reach in healthcare-underserved regions. Addressing this disparity, the present investigation proposes a compact, fully automated grading pipeline built around a lightweight deep architecture designed without reliance on externally pretrained weights. The framework sequences four tightly coupled stages—photometric correction, class-aware augmentation, hierarchical feature distillation, and discriminative dimensionality reduction—before assigning one of five internationally recognised severity labels through a probabilistic output layer. Systematic evaluation on two open-access fundus repositories produced a best-epoch test accuracy of 74.3%, a macro-averaged AUC-ROC of 0.83, and a weighted F1-score of 0.72, accompanied by transparent quantitative analysis of inter-epoch behavioural dynamics. These outcomes, alongside observed generalisation volatility, delineate both current capabilities and the roadmap toward clinical-grade deployment