Diabetic Retinopathy Classification using VisionTransformer Architectures and Deep Learning

Abstract Diabetic Retinopathy (DR) is a vision-threatening disease that has affected plenty of people worldwide and is treated using digital fundus imaging and retinal images by specialists at the primary level. DR is one of the most common eye diseases and it is caused by long-term diabetes and high blood pressure.A lot of research has been carried out in the recent past using Deep Learning (DL) models on eye disease diagnosis. However, recently Transformer based models have proved to be quite dynamic in terms of image classification and have outperformed DL architectures in many cases. In this research, we implement a vision Transformer (ViT) model with Shifted Patch Tokenization(SPT) and Locality Self Attention (LSA) that is specifically built to perform well on small datasets, alongside we implement someConvolutional Neural Network (CNN) based Transfer learning algorithms such as ResNet101, InceptionV3, and Xception. Upon training and testing the selected architectures, we find that theViT-based model outperforms all the state-of-the-art DL models ,attaining accuracies of 82.26% and 77.01% for training and validation respectively. Furthermore, we used StyleGAN2 as an alternative to the generic data augmentation technique to produce images that replicate the original properties of the real images.We augmented our dataset consisting of 2750 images usingStyleGAN2 and extended it into a total of 5000 images keeping an equal number of images in each of the classes mitigating class imbalance and ensuring the dataset is small enough to demonstrate ViT on small datasets.