GlaucoDiff: A Framework for Generating Balanced Glaucoma Fundus Images and Improving Diagnostic Performance

Abstract

Glaucoma is a leading cause of irreversible blindness, and early diagnosis is critical. While retinal fundus images are commonly used for screening, AI‐based diagnostic models face challenges such as data scarcity, class imbalance, and limited image diversity. To address this, we introduce GlaucoDiff, a diffusion‐based image synthesis framework designed to generate clinically meaningful glaucoma fundus images. It employs a two‐stage training strategy and integrates a multimodal large language model as an automated quality filter to ensure clinical relevance. Experiments on the JustRAIGS dataset show that GlaucoDiff outperforms commercial generators such as DALL‐E 3 and Keling, achieving better image quality and diversity (FID: 109.8; SWD: 222.2). When synthetic images were used to augment the training set of a vision transformer classifier, sensitivity improved consistently from 0.8182 with only real data to 0.8615 with 10% synthetic images, and further to 0.8788 with 50%. However, as the proportion of synthetic data increased, other important metrics such as specificity, accuracy, and AUC began to decline compared to the results with 10% synthetic data. This finding suggests that although more synthetic images can enhance the model's ability to detect positive cases, too much synthetic data may reduce overall classification performance. These results demonstrate the practical value of GlaucoDiff in alleviating data imbalance and improving diagnostic accuracy for AI‐assisted glaucoma screening.