A Lesion-Based Diabetic Retinopathy Detection Through Hybrid Deep Learning Model

Diabetic retinopathy (DR) can be defined as visual impairment caused by prolonged diabetes affecting the blood vessels in the retina. Globally, it stands as the primary contributor to blindness, impacting approximately 191 million individuals. While prior research has addressed DR classification using retinal fundus images, existing methods often focus on isolated lesion detection, lacking a comprehensive framework for the simultaneous identification of all lesions. Previous studies concentrated on early-stage features like exudates, aneurysms, hemorrhages, and blood vessels, sidelining severe-stage lesions such as cotton wool spots, venous beading, very severe intraretinal microvascular abnormalities (IRMA), diffuse intraretinal hemorrhages, capillary degeneration, highly activated microglia, and retinal pigment epithelium (RPE) damage. In this study, a deep learning approach is proposed to classify DR fundus images by severity levels, utilizing GoogleNet and ResNet models based on adaptive particle swarm optimizer (APSO), for enhanced feature extraction. The extracted features from the hybrid model are further used with different machine learning models like random forest, support vector machine, decision tree, and linear regression models. Experimental results showcased the proposed hybrid framework outperforming advanced approaches with a remarkable 94% accuracy on the benchmark dataset. This method demonstrates potential enhancements in precision, recall, accuracy, and F1 score for different DR severity levels.