<![CDATA[Retinal diseases such as Diabetic Retinopathy, glaucoma, and cataracts are major causes of global blindness that require early detection. However, manual screening faces challenges related to scalability and consistency, especially in resource-limited settings. This research aims to evaluate the comparative performance of six state-of-the-art Convolutional Neural Network (CNN) architectures to identify the most optimal model for the four-class retinal disease classification (cataract, diabetic retinopathy, glaucoma, and normal). The five models evaluated are EfficientNetB0, InceptionV3, MobileNetV2, VGG16, and VGG19. All models were trained using a transfer learning approach on the "eye_diseases_classification" dataset, compiled from various sources to ensure model generalization. The results of the comparative evaluation show that all models successfully achieved an accuracy above 90%, confirming the effectiveness of transfer learning in this task. However, the EfficientNetB0 architecture demonstrated the best performance, recording the highest accuracy of 93.67% and the best balance in precision and F1-Score metrics. The EfficientNetB0 model is established as the most reliable solution and is recommended for the development of an efficient automated early detection system to support clinical workflows.]]>
Copyrights © 2026
