Published Paper Details:

Diabetes mellitus (DM) represents a significant global health challenge, impacting 537 million adults across the globe. Among its various complications, diabetic retinopathy (DR) stands out as the primary cause of preventable blindness. The traditional manual screening for DR is costly, invasive, and reliant on skilled ophthalmologists, rendering it impractical for widespread screening in areas with limited resources. This research introduces a deep learning-driven automated framework for DR detection utilizing retinal fundus images. We assessed DenseNet121, EfficientNet-B4, and an Ensemble model that integrates both architectures using the EyePACS dataset. The preprocessing steps included green-channel extraction, Contrast Limited Adaptive Histogram Equalization (CLAHE), and comprehensive data augmentation to enhance lesion visibility and address class im- balance. The Ensemble model achieved an accuracy of and a Cohen's Kappa of surpassing earlier techniques, including the Symmetric CNN (SCNN) method. This study highlights the promise of AI-enhanced, non-invasive retinal screening for the early detection of DR and the facilitation of telemedicine-based healthcare services.