Deep Learning-Based Detection of Diabetic Retinopathy: A Comparative Study of CNNs and Vision Transformers

This Diabetic Retinopathy is a condition which happens when there is a sudden increase in the blood sugar level of vessels which may lead to vision related problems. In this study we have tested the performance of CNN, ViT and their hybrid models. The dataset used is publicly available on Kaggle and the dataset contained around 35,000 retinal images which were divided into 5 classes namely No DR, Mild DR, Moderate DR, Severe DR and Proliferative DR. In CNN we tested 4 different architectures in which we achieved the best accuracy of 75.4% with Resnet50 architecture and with ViT model we achieved an accuracy of 83.9% and from the hybrid model we achieved an accuracy of 88.4% from the Resnet50 + ViT. The results shown by the models were promising but there were some gaps in the study. The dataset used was skewed towards NO DR class. For future work more balanced datasets with some data augmentation techniques could be used. Additionally, the study used only 50 epochs which can be increased in future work to use the model to their full potential.