Diffusion‑synthesized Chest X‑rays improve fairness and diagnostic performance

Abstract

Deep learning models have been widely applied to chest X-ray (CXR) disease classification and diagnosis; however, challenges such as data scarcity and shortcut learning often lead to biased model behavior. This study addresses fairness-related concerns in conventional deep learning models trained on CXR data and proposes mitigating demographic disparities through image synthesis. We fine-tune a pre-trained stable diffusion model using Low-Rank Adaptation (LoRA) and a CLIP tokenizer, incorporating low-rank constraints into key attention layers while preserving the original architecture. This enables the generation of high-quality, realistic CXR images with reduced parameter complexity. Experimental results demonstrate that models trained with our synthetic data achieve improved classification performance and exhibit significantly reduced disparities across demographic groups. Furthermore, the proposed models show increased attention to disease-relevant regions and diminished reliance on spurious shortcuts. These findings highlight the potential of generative AI in enhancing fairness in medical imaging workflows, particularly when combined with efficient and adaptable fine-tuning strategies.