<![CDATA[Eye diseases serve as a primary contributor to global blindness, making early detection a critical determinant in effective treatment outcomes. While retinal fundus image analysis is the diagnostic standard, conventional manual methods are often hindered by observer subjectivity and time inefficiencies. This study aims to optimize eye disease classification using a Convolutional Neural Network (CNN) approach empowered by transfer learning techniques. Utilizing a dataset of 1,200 retinal fundus images sourced from Kaggle, this research classifies four categories: normal, glaucoma, cataract, and diabetic retinopathy. To mitigate the challenge of limited labeled medical datasets, specific data augmentation strategies—including random flip, zoom, and contrast adjustments—were applied. The study conducts a comparative evaluation of three architectures: standard VGG16, baseline MobileNet, and a proposed optimized MobileNet. The proposed method utilizes Random Search to systematically optimize hyperparameters such as learning rates, dense layer units, and dropout rates. Experimental results demonstrate that the optimized MobileNet achieved superior performance with 89.17% accuracy, significantly outperforming the VGG16 baseline 82,00% and baseline MobileNet 85,00%. Notably, the model achieved perfect recall for diabetic retinopathy, although glaucoma remained the most challenging class due to subtle morphological similarities with normal eyes. These findings confirm that integrating lightweight CNNs with appropriate transfer learning yields a diagnostic system that is not only accurate but also efficient for deployment in resource-constrained environments]]>
Copyrights © 2026
