Enhancing early detection of diabetic retinopathy using vision transformer and classification models

This study enhances early detection of DR using Vision Transformer (ViT) and comparative classification models, PCA with CNN, and Dual CNN with Inception-v3 on multiple retinal image datasets such as REFUGE1-train, EyePACS-Glaucoma, OIA-ODIR-TRAIN, and G1020. After preprocessing and model optimization, the ViT model achieved the highest accuracy of 95%, followed by Dual CNN with Inception-v3 at 94%, and PCA with CNN at 93%. Performance metrics for ViT recorded a precision of 0.96, recall of 0.97, and F1-score of 0.95, confirming its superior diagnostic capability. The results demonstrate that the transformer's self-attention mechanism effectively captures global retinal features, outperforming convolution-based models. This work establishes ViT as a reliable and efficient deep learning framework for automated DR detection, supporting early clinical intervention and improved patient outcomes.