A Deep Learning-driven Approach for Early Detection of Ocular Abnormalities Using MobileNet and EfficientNet

Abstract

Diabetic retinopathy, hypertensive retinopathy, glaucoma, and cataract are well-established eye diseases resulting from elevated blood pressure, increased blood glucose levels, and heightened eye pressure. Symptoms typically manifest at later stages, encompassing phenomena such as AV (arteriovenous) nicking, constricted veins in the optic nerve, cotton wool patches and blood accumulation in the optic nerve. These pathologies can progress to severe complications, including retinal artery occlusion, optic nerve damage, and the potential for irreversible vision impairment. The integration of artificial intelligence (AI) and deep learning models offers a promising prospect for early disease detection. This study utilizes datasets sourced from reputable internet platforms to introduce a novel methodology called CAD-EYE designed for the classification of diabetic retinopathy, hypertensive retinopathy, glaucoma, and cataract. CAD-EYE employs MobileNet and EfficientNet models, with a particular emphasis on feature fusion to enhance overall performance of the diagnostic system. The system has been trained on 65,871 digital fundus images sourced from diverse datasets. In a comparative analysis, CAD-EYE outperforms state-of-the-art models such as CNN+LSTM, ResNet, GoogleNet, VGGNet, InceptionV3, and Xception in terms of classification accuracy. These findings underscore the efficacy of CAD-EYE as an adept diagnostic tool, designed not to replace optometrists but to complement the efforts of healthcare professionals by providing valuable assistance in the early identification of ocular pathologies.