A Trustworthy Hybrid Multibranch Architecture for Robust Diabetic Retinopathy Classification Under Data Scarcity

Diabetic retinopathy (DR) is a leading cause of preventable blindness worldwide, making early and accurate screening essential. However, automated DR grading from fundus images remains difficult due to insufficient expert annotations and severe class imbalance. To address these challenges, we propose a hybrid multi-branch deep learning framework that integrates a convolution–attention Hybrid Feature Extraction Block (HFEB) and a Multi-Branch Feature Fusion (MBFF) block. A pre-trained Vision Transformer (ViT) is used to adaptively weight branch features for dynamic integration, and a Parallel Feature Fusion module and composite loss function are applied to enhance lesion feature learning and mitigate class imbalance. Extensive experiments show that our method achieves state-of-the-art accuracies of 88.5%, 86.8%, and 73.4% on the APTOS2019, DDR, and IDRiD datasets, with strong generalization ability on a private ROP dataset. Grad-CAM visualization verifies that the model focuses on clinically important lesions. Furthermore, the framework is designed for edge deployment on IoT medical terminals. The lightweight version achieves real-time inference of 33.2ms on Jetson AGX Orin 64G. Even with ViT guidance, the inference latency is still within the acceptable range for hospital AI detection equipment, meeting the low-latency and low-cost demands of portable screening devices and supporting scalable telemedicine IoT applications.