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 represents a leading cause of preventable blindness. Automated screening systems face critical challenges: severe class imbalance in medical datasets where healthy cases vastly outnumber vision-threatening stages, and computational demands limiting deployment in resource-constrained settings. This research presents a comprehensive three-stage framework integrating Progressive Growing GANs for synthetic fundus image generation, a dual-branch ensemble 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, achieving FID score of 23.4. On APTOS 2019, the ensemble 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. The distilled MobileNetV2 student retains 90.0% accuracy and 0.938 kappa while reducing parameters by 96.8% and computational operations by 98.8%. 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.