A ResNet Transformer Hybrid Approach for Enhanced Detection of Retinal Diseases

This paper presents a hybrid deep learning model that combines ResNet50 and Transformer encoder layers for multiclass classification of retinal diseases from fundus images. Using a publicly available dataset, we demonstrate that our architecture outperforms standard CNN baselines, achieving a test accuracy of 94.20%. The robustness of the model to class imbalance is enhanced by a boundary-aware focal loss. The proposed model shows consistent performance across all four categories of retinal diseases, with particularly high precision and recall for diabetic retinopathy and cataract. Notably, it achievesa glaucoma recall of 91.91%, addressing a common weakness in many existing models. Validation and testing metrics indicate strong generalization and minimal overfitting, supported by stable learning curves and class-wise evaluation. The model also demonstrates resilience to challenging cases near decision boundaries, improving diagnostic reliability in real-world scenarios. The proposed approach demonstrates that hybrid architectures can significantly enhance automated retinal disease screening by achieving high overall accuracy and balanced performance across multiple disease classes. This makes it a promising tool for sup- porting ophthalmologists in the early detection and classification of retinal disorders in real-world screening environments.