Comparative Analysis Of Deep Learning Architectures For The Classification Of Eye-Related Diseases

Eye-related disorders are among the driving causes of vision impedances worldwide. Accurate classification of retinal conditions is crucial for adequate treatment and management. This research explores the potential of three advanced deep-learning models, namely—-VGG16, DenseNet, and Swin Transformers, to classify four eye-related disorders utilizing the Optical Coherence Tomography Image Retinal Database (OCTID) taken from ICPSR. The OCTID dataset contains retinal OCT images representing four distinct pathological conditions: age-related macular degeneration, diabetic retinopathy, central serous retinopathy, macular holes, and normal retinas. Evaluation is done based on metrics such as accuracy, precision, recall, and F1 score. The comparison of these models, paired with interpretability techniques such as GRADCAM and Integrated Gradients, reveals that VGG16 achieves the highest F1-score, demonstrating robust classification performance. At the same time, Swin Transformers develops competitive accuracy with practical interpretability through Integrated Gradients. DenseNet showed slightly lower metrics but performed adequately on the GRADCAM outputs. This study underscores the critical role of advanced deep learning frames, which work to enhance diagnostic accuracy for retinal disorders and provide a foundation for future developments in ophthalmic disease classification.