Advanced retinal disease detection using RHT-Net: a hybrid deep learning approach with augmented fundus imaging.

Introduction: Retinal diseases are a leading cause of visual impairment worldwide, highlighting the need for accurate and scalable automated screening systems. This study introduces RHT-Net (Retinal Hybrid Transformer Network), a novel hybrid deep learning architecture designed for multi-class classification of nine retinal diseases from color fundus images by combining the strengths of convolutional neural networks and transformer encoders. Methods: RHT-Net integrates residual convolutional neural networks for local feature extraction with transformer encoders to capture long-range global dependencies. A real-world dataset comprising 5,318 color fundus images collected from Bengali patients was used in this study. Data augmentation techniques, including rotation, flipping, and Gaussian noise addition, expanded the dataset to 21,272 images. Images were preprocessed through resizing to 224 × 224 pixels and contrast enhancement using Contrast Limited Adaptive Histogram Equalization (CLAHE). The augmented dataset was divided into training (80%) and testing (20%) subsets for model development and evaluation. Results: The proposed RHT-Net achieved a training accuracy of 97.93% with an F1-score of 96.10%. On the testing set, the model attained an accuracy of 96.12% and an F1-score of 92.28%. The overall classification performance reached an accuracy of 97.57% and an F1-score of 95.31%. Comprehensive evaluations, including class-wise performance analysis, confusion matrices, and Grad-CAM visualizations, demonstrated the model's strong predictive capability, robustness, and interpretability across the nine retinal disease classes. Discussion: The findings indicate that RHT-Net is a promising and scalable approach for early retinal disease screening. By effectively capturing both local and global image features, the model achieves high classification performance while providing interpretable predictions. These characteristics support its potential integration into telemedicine and remote diagnostic workflows. However, further external validation and deployment-oriented optimization are necessary before real-world clinical implementation.