An efficient deep learning framework for diabetic retinopathy classification using generative data augmentation and knowledge distillation

Abstract

Diabetic retinopathy (DR) affects 93 million people globally and is a leading cause of preventable blindness. Automated screening systems face two critical barriers: severe class imbalance in medical datasets where healthy cases vastly outnumber vision-threatening stages, and high computational demands that limit deployment in resource-constrained settings. This research presents a three-stage framework integrating Progressive Growing GANs for synthetic fundus image generation, a dualbranch ensemble architecture combining EfficientNet-V2-S with Vision Transformer- B/16, and knowledge distillation to compress the model into MobileNetV2. The generative approach created a balanced training corpus of 10,000 images across five severity classes, addressing the fundamental imbalance problem while maintaining pathological authenticity validated through perceptual metrics. On APTOS 2019, the ensemble teacher achieves 95.8% accuracy with 0.975 quadratic weighted kappa. Cross-dataset validation on DDR and Messidor-2 confirms robust generalization across diverse clinical populations and imaging protocols. The distilled MobileNetV2 student retains 92.4% accuracy and 0.938 kappa while reducing model parameters by over ninety-six percent and computational operations by over ninetyseven percent. Inference latency decreases by over eighty percent on CPU, with model size compressed to enable mobile deployment. Grad-CAM visualizations confirm clinically relevant attention to microaneurysms, hemorrhages, and neovascularization. This framework demonstrates that generative augmentation combined with heterogeneous ensemble learning and knowledge distillation overcomes the accuracydeployability trade-off, enabling accessible DR screening in resource-limited settings.