Glaucoma is the leading cause of irreversible blindness worldwide. Early detection is essential to preserve vision. Deep learning approaches have shown promise in automating glaucoma detection. However, significant class imbalance in medical datasets often impairs classifier performance. To address this challenge, we propose AUBADE-syn, a deep learning ensemble framework that integrates synthetic image generation with structured class-balancing strategies. Our approach leverages optic nerve head-centered regions and a classifier-free guided diffusion model to generate realistic glaucomatous images, enriching the minority class and improving model generalization on highly imbalanced datasets. We benchmarked the AUBADE-syn algorithm against widely used methods for addressing class imbalance, including weighted loss functions, focal loss, Balanced-MixUp, ProCo, and FlexDA. On EyePACS, a large-scale public dataset with a 1:30 class imbalance ratio, AUBADE-syn achieved an area under the receiver operating characteristic curve of 0.992, outperforming all comparison methods. We also validated its performance across ten independent public datasets and fine-tuned the model on three additional public datasets, achieving top-tier or competitive results relative to previously published methods. Overall, these results demonstrate that AUBADE-syn consistently improves both discrimination and calibration for glaucoma detection in highly imbalanced settings, highlighting the effectiveness of domain-aware synthetic augmentation and structured ensemble learning for imbalanced medical imaging tasks.
Wu et al. (2026) studied this question.