Transformer-Based Multi-Class Model for Diabetic Retinopathy Using DeiT and ViT on Fundus Images

Diabetic retinopathy (DR) is one of the leading causes of preventable vision loss worldwide, mainly due to challenges in early and accurate diagnosis. Traditional diagnostic methods depend on expert evaluation, which can be time-consuming and subjective. This study proposes a transformer-based multi-class classification framework for DR severity detection, leveraging two models, Data-Efficient Image Transformer (DeiT) and Vision Transformer (ViT). Utilizing the DDR dataset comprising five distinct DR severity levels, we address class imbalance through comprehensive data augmentation and balancing strategies. Experimental results demonstrate competitive performance, with the DeiT model achieving a test accuracy of 91.07 % and the ViT model 90.98 %. DeiT exhibits superior adaptability and robustness in capturing subtle retinal features critical for early-stage DR detection. This work highlights the potential of transformer architectures as efficient and scalable tools for automated DR screening, paving the way for improved clinical decision support in diabetic care.