COMPARATIVE ANALYSIS OF VISION TRANSFORMERS AND CONVOLUTIONAL NEURAL NETWORKS IN DIABETIC RETINOPATHY DIAGNOSIS

Abstract

Diabetic retinopathy can lead to significant visual complications and significantly affects individuals' quality of life. This study focuses on comparing the performance of Vision Transformer (ViT) models and Convolutional Neural Networks (CNN) methods in diabetic retinopathy diagnosis and aims to evaluate their potential as an alternative to traditional diagnostic methods. In this study, the performance of four different ViT model architectures and four different convolutional neural network (CNN) models in training and testing phases were comparatively analyzed. ViT models achieved accuracy rates of 97.83%, 98.41%, 95.2%, and 98.26% for \"tiny,\" \"base,\" \"small,\" and \"large,\" respectively. Additionally, models trained with VGG13, ResNet18, ResNet50, and SqueezeNet architectures from CNN techniques achieved accuracy rates of 96.1%, 97.83%, 90.9%, and 93.93%, respectively. ViT architectures achieved higher accuracy rates than CNN architectures. When the results were evaluated, it was concluded that ViT methods were more successful in the diagnosis of diabetic retinopathy.