Diffusion-synthesized Chest X-rays improve fairness and diagnostic performance.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models for chest X-ray (CXR) analysis face challenges like data scarcity and shortcut learning, leading to biased outcomes.
• Existing models often exhibit demographic disparities in disease classification and diagnosis.

Purpose of the Study:

• To address fairness concerns in CXR deep learning models.
• To mitigate demographic disparities using synthetic image generation.
• To improve model performance and reduce reliance on spurious correlations.

Main Methods:

• Fine-tuned a pre-trained stable diffusion model using Low-Rank Adaptation (LoRA) and a CLIP tokenizer.
• Incorporated low-rank constraints into attention layers to generate high-quality, realistic CXR images.
• Trained classification models on both real and synthetic CXR data.

Main Results:

• Models trained with synthetic data achieved improved classification performance.
• Significantly reduced disparities across demographic groups were observed.
• Models showed increased attention to disease-relevant regions and diminished reliance on shortcuts.

Conclusions:

• Generative AI, combined with efficient fine-tuning, can enhance fairness in medical imaging.
• Synthetic data generation is a viable strategy to combat bias in AI-driven CXR analysis.
• The proposed method offers a promising approach for developing more equitable medical AI tools.