Dual Attention Based Dual Vision Transformer (DAD-ViT) for Diabetic Retinopathy Severity Grading Using Fundus Images

The study presents a Dual Attention Based Dual Vision Transformer model which researchers tested with fundus images to assess Diabetic Retinopathy severity. The system design uses two different transformer branches which process visual data through distinct patch dimensions of $16 \times 16$ and $32 \times 32$ to capture different visual information. The system uses combined channel and spatial attention methods to show important medical information. The DAD-ViT model achieved 98.72 % accuracy with 98.11 % precision and 97.94 % recall and 98.02 % F1-score and 0.994 AUC and 0.963 Cohen's Kappa rating when evaluated on the APTOS 2019 and Messidor datasets which contained 4862 images through five-fold cross-validation testing. The results show substantial improvement beyond normal human observer matching results which show inter-observer agreement between 0.85 and 0.92 Cohen's Kappa. The system develops new capabilities through dual-scale processing and integrated attention architecture which outperforms both convolutional systems and single-branch transformers.