Deep Learning-Based Diabetic Retinopathy Classification Using Vision Transformers (VIT): A Comparative Study

The paper will elaborate on the different models of deep learning that are used to detect and classify Diabetic Retinopathy using retinal fundus images. Diabetic Retinopathy is one of the life-threatening complications of diabetes and one which limits vision and has affected millions of people across the world, creating serious personal, social, and economic problems. Early diagnosis and successful treatment of DR are highly significant in avoiding blindness and enhancing the quality of life in patients. This paper is dedicated to the most recent techniques of Diabetic Retinopathy analysis based on Vision Transformers architecture deployed on PyTorch. We apply a strong attention algorithm of transformer for extracting global and local details in retinal images with the aim of achieving better diagnostic results. Furthermore, this paper determines ViT performance compared to other deep learning models, including CNNs, in order to establish the efficiency of each model in medical image classification. In addition, the paper discusses contemporary issues of DR detection including imbalance in data, interpretability, and cross-population generalization of the model. Finally, the future directions discussed are the integration of multi-modal data sources, the use of explainable artificial intelligence (XAI) to promote transparency, and mobile-friendly diagnostic tools that could enable accessible and real-time screening of diabetic retinopathy.