Advancing Diabetic Retinopathy Screening With DR-NetFusion: A Hybrid Deep Learning Model for Enhanced Detection and Interpretability

Diabetic retinopathy (DR) is one of the leading causes of preventable blindness worldwide, making timely and accurate detection crucial for effective management. This study introduces DR-NetFusion, a novel hybrid deep learning framework designed to automate DR detection and classification. The proposed model synergistically combines convolutional neural networks (CNNs) and transformer architectures, leveraging the strengths of both in capturing local features and global context from retinal images. DR-NetFusion performs multiscale feature extraction, integrates a dual-attention mechanism, and incorporates ensemble learning to improve robustness and model performance. Additionally, the framework utilizes generative adversarial networks (GANs) for synthetic data augmentation to address data scarcity challenges and applies pretrained transfer learning to enhance efficiency. For interpretability, we incorporate Grad-CAM and SHAP techniques, providing visualizations that improve clinical trust. Extensive evaluations on large-scale datasets, including Kaggle EyePACS, Messidor, and IDRiD, demonstrate that DR-NetFusion achieves state-of-the-art results with sensitivities of 97.8%, specificities of 96.7%, and a weighted F1-score of 0.93 for DR grading. This research presents a comprehensive and highly accurate solution for DR screening, offering significant potential for early diagnosis and improved treatment strategies in ophthalmology.