<p>Diabetic Retinopathy is one of the major causes of vision impairment worldwide, and its timely diagnosis can prevent blindness. In general, traditional centralized deep learning methods are faced with challenges relating to privacy issues, the sharing of limited data, and non-uniform distributions. The authors have proposed a Federated Vision Transformer framework, FedViT, for a privacy-aware collaborative DR detection across various institutions without necessarily sharing raw patient data. FedViT integrates the Federated Averaging algorithm with Adam Weight Decay optimization for its robust convergence to a superior generalization. Experimental results on the Kaggle and APTOS datasets show that FedViT outperforms state-of-the-art architectures such as ResNet101, VGG19, and VGG16 by achieving 99.02% and 97.7% accuracy, respectively. Increasing client participation and local training rounds improves accuracy while reducing communication overhead, thus validating the scalability and efficiency of the proposed framework. It therefore indicates the importance of federated learning in medical domains by providing an efficient, secure, and high-performance solution for real diabetic retinopathy detection and any other sensitive healthcare applications.</p>