Deep Learning Based Classification of Diabetic Retinopathy and Glaucoma for Early Comorbidity Detection

Abstract

Vision-threatening retinal diseases such as Diabetic Retinopathy (DR) and newlineGlaucoma continue to be leading causes of irreversible blindness globally, particularly newlinein individuals with diabetes and aging populations. Traditional screening methods rely newlineheavily on expert analysis of fundus images, which are often time-consuming, newlinesubjective, and inconsistent. This thesis introduces three successive, deep learning-based newlineresearch contributions to address these challenges with enhanced diagnostic accuracy, newlinerobustness, and clinical applicability. newlineThe first contribution introduces a comprehensive pipeline MHSAGGCNBCOA, newlineThis method integrates intelligent image enhancement, advanced segmentation, newlineand a deep graph-based classifier. Abnormalities like microaneurysms and exudates are newlinethen segmented using Federated Fuzzy K-Means Clustering. Feature extraction is newlineperformed using a Force-Invariant Improved Feature Extraction method that prioritizes newlinecontrast, energy, correlation, and homogeneity metrics. To reduce dimensionality and newlineredundancy, the Binary Chimp Optimization Algorithm (BCOA) selects optimal newlinefeatures. The core classifier, a Multi-Head Self-Attention Gated Graph Convolutional newlineNetwork (MHSAGGCN), captures spatial relationships in the extracted features, newlineensuring precise staging of DR into Normal, Earlier, Moderate, and Severe classes. This newlinemodel achieved 99.5% accuracy on the DIARETDB1 dataset newline