Enhanced DR Screening Using DiabDHS-Net: A Deep Learning Approach with FGK Clustering and Morphological Processing

Diabetic Retinopathy (DR) develops as a microvascular diabetes complication that affects the retinal blood vessels and causes vision loss that could become permanent unless diagnosed and treated appropriately. Research on deep learning-based detection of DR has gained extensive popularity due to improvements in medical image processing and neural network designs. The DiabDHS-Net model uses DenseNet201 with the Human Evolutionary Optimization Algorithm (HEOA) and Softmax classifier for automated DR detection. The novelty of this model is that DenseNet extracts dense features with maximum information retention, and HEOA optimizes networks to find distant global solutions suitable for medical applications. The identification process for the fovea region depends on FGK Clustering (Fuzzy Gustafson-Kessel) to achieve high accuracy results. Experimental investigation in the Messidor dataset demonstrates that the proposed model surpasses existing models by achieving 99.53% accuracy, 99.12% precision, 99.67% recall, and 98.9% F1-score, along with 99.81% specificity. ROC curve and feature importance plots prove that the model distinguishes different DR severity levels (No DR, Mild DR, Moderate DR, Severe DR) with high levels of sensitivity and specificity.