Integrating Deep Learning Techniques for Automated Retinal Disease Detection Using Fundus Images

Abstract

RetinalFormer was compared with eight state-of-the-art models, including pure CNN architectures (VGG-19, ResNet-50, DenseNet-121, EfficientNetV2-M), pure Transformer models (ViT-Base/16, Swin-Transformer-B), and existing hybrid CNN-Global attention models. Globally, nearly 2.2 billion people have visual impairments, with at least one billion due to avoidable or treatable causes. Retinal disorders, such as Diabetic Retinopathy (DR), Glaucoma, and Age-related Macular Degeneration (AMD), contribute significantly to avoidable blindness but tend to go unnoticed due to their asymptomatic nature until irreversible stages occur. Fundus photography offers a simple, inexpensive approach to screening patients’ retinas; however, it has limitations due to a shortage of ophthalmologists and the need for subjective interpretation.

This paper proposes an effective deep learning architecture for automated and multi-class classification of retinal diseases from fundus images. Our proposed model includes a transfer-learning-based Image Quality Evaluation Tool (QET) to filter out low-quality images. Contrast enhancement using the CLAHE method and class-specific data augmentation are considered to address class imbalance. We compare three state-of-the-art deep learning architectures: ResNet-152, EfficientNetV2, and YOLOv11. The multi-Scale Attention Transformer (MSAT) and the Hybrid DenseNet-VGG16 model are used for specific tasks. Feature optimization is performed using the Bitterling Fish Optimization (BFO) algorithm, while hyperparameter tuning is performed using the Honey Badger Optimization (HBO). Explainability is ensured through Grad-CAM heatmaps and t-SNE visualizations. The proposed model is evaluated on standard benchmarking datasets, including RFMiD, ODIR-5K, Drishti-GS1, RIM-ONE, and ORIGA-Light, achieving classification accuracies above 90% compared to existing state-of-the-art models.