Abstract

This study evaluates the effectiveness of Vision Transformers (ViTs) and hybrid deep learning architectures for diabetic retinopathy (DR) classification, addressing the challenge of inter-stage ambiguity in traditional systems. While convolutional neural networks (CNNs) such as ResNet50 excel at localized feature extraction in retinal images, ViTs offer superior global contextual modeling. To synergize these strengths, we propose a hybrid architecture integrating ResNet50’s granular feature extraction with ViTs’ global relational reasoning. Three models are designed and evaluated: (1) an auto-tuned ResNet50, (2) a hyperparameter-optimized ViT, and (3) a hybrid combining both architectures.
To reduce ambiguity between neighboring stages, we simplified the traditional five-stage classification into three clinically relevant categories: no DR, early DR (mild/moderate), and advanced DR (severe/proliferative).
Trained and validated on the APTOS dataset. The ResNet50 model achieves precision scores of 93.0% (No DR), 82.0% (Early DR), and 86.0% (Advanced DR). The standalone ViT demonstrates relative improvements, attaining 98.0%, 91.0%, and 93.0%, respectively. The hybrid model surpasses both, achieving 98.0% average precision across all classes, with gains of +7.0% (early DR) and +5.0% (advanced DR) over the standalone ViT. The proposed hybrid model achieved an impressive value of 99.5% on all metrics (accuracy, precision and recall) for identifying DR (binary classification) and a value of 98.3% for 3-stage classification. It was also concluded that the proposed method achieved better performance in DR detection and classification compared to conventional CNN and other state-of-the-art methods}.The proposed hybrid approach significantly reduces confusion between classes, demonstrating its potential for accurate classification of the different stages of DR.

Key words: Retinopathy Diabetic, Vision transformer, Transfer Learning, Artificial Intelligence