Abstract
Diabetic retinopathy (DR) is a common diabetes complication that presents significant diagnostic challenges due to its reliance on expert assessment and the subtlety of small lesions. Although deep learning has shown promise, its effectiveness is often limited by low-quality data and small sample sizes. To address these issues, we propose a novel deep learning framework for DR that incorporates self-paced progressive learning, introducing training samples from simple to complex, and randomized multi-scale image reconstruction for enhanced data augmentation and feature extraction. Additionally, ensemble learning with Kullback-Leibler (KL) divergence-based collaborative regularization improves classification consistency. The method's effectiveness is demonstrated through experiments on the integrated APTOS and MESSIDOR-Kaggle dataset, achieving an AUC of 0.9907 in 4-class classification, marking a 2.2% improvement compared to the ResNet-50 baseline. Notably, the framework achieves a recall of 97.65% and precision of 96.54% for the No-DR class, and a recall of 98.55% for the Severe class, with precision exceeding 91% across all categories. Furthermore, superior classification performance on limited data samples, as well as robust localization of subtle lesions via multi-scale progressive learning, has been demonstrated, underscoring the potential of the proposed framework for practical clinical deployment.