Vision Transformer for Diabetic Retinopathy Diagnosis: A Deep Learning Approach Beyond CNN

Automated retinal disease detection, particularly in Diabetic Retinopathy (DR) classification, has received significant attention. While Convolutional Neural Networks (CNNs) have been the standard for medical image classification, the Vision Transformer (ViT) has emerged as a promising alternative. This study compares ViT and CNN models on the APTOS 2019 Blindness Detection dataset, evaluating accuracy, precision, recall, F1-score, ROC-AUC, and computational efficiency. ViT achieved the highest accuracy (99.3%) and ROC-AUC (0.999), demonstrating superior precision and recall for both Normal and DR classes but required significantly higher computational resources. In contrast, CNN models such as EfficientNet-B0 (98.36% accuracy, ROC-AUC 0.9837) and VGG16 (ROC-AUC 0.9836) were more computationally efficient. DenseNet121 and ResNet18 followed with slightly lower accuracies. This study highlights the trade-off between performance and computational cost, offering insights into the feasibility of ViT in real-world medical applications.