Benchmarking CNN Architectures for Eye Disease Detection with Transfer Learning Techniques

In this study, convolutional neural networks (CNN)-based approaches were compared to classify eye diseases using transfer learning techniques. A series of data augmentation strategies, including random rotation, shifting, shearing, zooming, and horizontal flipping, were applied to increase the training data’s robustness and diversity. Several state-of-the-art CNNs, including ResNet50, VGG19, EfficientNetB0, Xception, InceptionV3, DenseNet121, MobileNetV2, NASNetMobile, and ConvNeXtBase, were fine-tuned through transfer learning. During training, models were evaluated based on their accuracy, training time, and validation performance, while early stopping mechanisms were employed to prevent overfitting. Experimental results demonstrated that DenseNet121 achieved the highest validation accuracy ($72 \%$) during the training phase and the best test set performance with an accuracy of $68 \%$ and an AUC-ROC of 0.93. MobileNetV2, on the other hand, provided a strong balance between classification accuracy ($65 \%$) and low inference time (7.28 ms), making it appropriate for real-time uses. The findings highlight the importance of selecting appropriate architectures by considering both predictive performance and computational efficiency, particularly in the context of medical imaging, where real-world deployment constraints are critical.