Abstract
Visual impairment is one of the leading causes of blindness. Optical Coherence Tomography (OCT) is an essential imaging technology to aid in the diagnosis of various retinal diseases. In this paper, the researchers have utilized Convolutional Neural Networks (CNNs) to provide accurate diagnosis for retinal diseases based on OCT images. The researchers compared the performance of three pre-trained CNN models-ResNet50, GoogLeNet, and MobileNetV2-each of which was trained using three optimizers: Adam, RMSProp, and Stochastic Gradient Descent with Momentum (SGDM). The networks were validated using the OCT2017 dataset, consisting of grayscale scans marked as CNV, DME, Drusen, or Normal. The most optimal model was GoogLeNet with RMSProp, which had a test accuracy of 97.26% and a superior generalization to the other tested models. GradCAM visualizations verified that the models were attending to clinically relevant retinal features, improving interpretability and facilitating clinical usability. These findings highlight modeloptimizer pairing relevance in medical image classification and show promising directions towards the application of deep learning systems in ophthalmic diagnosis.