Addressing Class Imbalance in Diabetic Retinopathy Detection: An Enhanced Diffusion-Resampling Data Augmentation Approach

Diabetic retinopathy (DR) is a severe complication of diabetes that can lead to vision impairment or blindness if not detected early. Machine learning (ML) models offer a promising approach for early DR detection by leveraging electronic health records (EHRs). However, class imbalance, where non-DR cases vastly outnumber DR cases, poses a significant challenge, often leading to inaccurate predictions. This study proposes a new data augmentation framework termed DDPM-COFFRe that integrates an enhanced denoising diffusion probabilistic model (DDPM) developed in this work, with resampling to address the data imbalance issue more effectively. The DDPM-COFFRe integrates cosine noise schedule, optimized contrastive learning, fairness metrics, and outlier filtering to generate high-quality synthetic DR samples. Moreover, it also involved ADASYN-based resampling to further refine data augmentation by targeting underrepresented regions of the minority class. The effectiveness of DDPM-COFFRe is evaluated using multiple classifiers across three test scenarios: an imbalanced test set, a weighted F1 evaluation, and a balanced test set. The results show that while augmentation improves recall and DR sensitivity, especially under balanced test conditions, its impact on overall classification performance remains limited in imbalanced, real-world scenarios. The findings suggest that while augmentation helps models identify DR cases, improvements are constrained by classifier limitations and the complexity of distinguishing DR from non-DR in imbalanced test sets.