Vision Transformer Approaches for Diabetic Retinopathy

The focus of this paper is to identify how deep learning can be utilised in identifying diabetic retinopathy, a serious complication of diabetes which causes blindness. In this paper, an end-to-end system is proposed for diagnosing the early signs of disease using CNN to analyze images. Cropped Retinal - CNN - Early Diagnosis of the First Stage of the Disease. The overall performance of the proposed system is determined using accuracy, precision, and usability that are far much better than the conventional method of diagnosing the syndromes. Five models are explored under this study which includes Vision Transformer (ViT), Convolutional Neural Network, MobileNetV2, YOLO and Hybrid model (ViT+CNN). five levels of Diabetic Retinopathy (DR) are explored viz., No, Mild, Moderate, Severe and Proliferative DR. From these five models, the hybrid model achieved a $\mathbf{9 4 \%}$ of accuracy on the DR dataset. More future work includes the expansion of the expressed model and the inclusion of real-time image processing for clinical use.