Comparative Study of Deep Learning Architectures for Automated Diabetic Retinopathy Grading: Vision Transformer, Swin Transformer, and InceptionResNetV2

Abstract

Diabetic Retinopathy (DR) is a vision-threatening complication of diabetes mellitus that progresses silently through five clinically defined severity grades. Timely automated screening is critical to prevent irreversible vision loss, particularly in resource-constrained healthcare settings. This paper presents a systematic comparative study of three state-ofthe-art deep learning architectures-Vision Transformer (ViTBase/16), Swin Transformer (swin base patch4 window7 224), and InceptionResNetV2-applied to five-class DR grading on the APTOS 2019 fundus image dataset (3,662 images). All models employ transfer learning from ImageNet-pretrained weights. We analyze each architecture from the perspectives of classification accuracy, per-class F1-score, macro-averaged AUC, GradCAMbased explainability, training dynamics, and parameter efficiency. Our ViT-Base/16 model, fine-tuned end-to-end with AdamW, cosine annealing, and label smoothing, achieves the highest validation accuracy of 85.40% with a macro-averaged F1-score of 0.7247. Swin Transformer achieves 83.20% accuracy, while InceptionResNetV2 achieves 81.40% through two-stage transfer learning. GradCAM visualizations confirm clinically aligned lesion localization across all architectures. This work provides architectural insights for deploying robust DR screening systems in clinical environments.