Enhancing Retinal Disease Detection With the Swin Transformer: A Comprehensive Comparative Analysis

ABSTRACT Retinal fundus image classification is paramount for the early detection and management of blinding eye diseases, such as diabetic retinopathy, glaucoma, and cataracts. This study systematically evaluates and compares various deep learning models for retinal disease classification, including Convolutional neural networks (CNNs), attention‐augmented CNNs, CNNs combined with long short‐term memory (LSTM) and attention, Vision Transformers (ViTs), and the Swin Transformer. These models were trained and tested on a large‐scale fundus image data set encompassing multiple retinal disease classes. Our experimental results demonstrate that the Swin Transformer model achieves the highest classification accuracy of 92%, outperforming both conventional CNN‐based models and the ViT. The Swin Transformer's shifted window attention mechanism facilitates efficient feature extraction, reducing computational complexity while maintaining high accuracy. We further evaluate the models based on performance metrics such as training time, GPU memory usage, and model size, providing a holistic assessment of their feasibility for real‐world applications. This study underscores the merits of transformer‐based architectures in retinal image analysis, particularly in enhancing classification accuracy while optimizing computational efficiency. The findings indicate that the Swin Transformer is a promising model for clinical applications, offering a robust and scalable solution for automated retinal disease detection. Notably, it achieved the highest classification accuracy of 92%, outperforming all other evaluated models.