Generative models improve fairness of medical classifiers under distribution shifts

Area of Science:

• Medical Imaging
• Machine Learning
• Generative Artificial Intelligence

Background:

• Domain generalization is a key challenge in healthcare machine learning, where models underperform due to data discrepancies between development and deployment.
• Underrepresentation of specific groups or conditions in training data leads to reduced model performance and fairness.
• Acquiring and labeling extensive clinical data is often infeasible due to cost and rarity of conditions.

Purpose of the Study:

• To investigate the use of generative artificial intelligence, specifically diffusion models, to create synthetic data for improving machine learning model robustness and fairness.
• To address the unmet need for label-efficient data augmentation in medical machine learning.
• To evaluate the effectiveness of learned augmentations across diverse medical imaging tasks.

Main Methods:

• Utilized diffusion models to learn realistic data augmentations in a label-efficient manner.
• Enriched training datasets with synthetic examples to address underrepresented conditions and subgroups.
• Evaluated model performance and fairness on three distinct medical imaging datasets: histopathology, chest X-ray, and dermatology images.

Main Results:

• Learned augmentations generated by diffusion models significantly improved model robustness across all tested medical imaging tasks.
• The approach enhanced statistical fairness, particularly improving diagnostic accuracy for underrepresented groups, especially in out-of-distribution settings.
• Synthetic data augmentation proved effective in histopathology, chest X-ray, and dermatology image analysis.

Conclusions:

• Generative AI, through diffusion models, offers a steerable and label-efficient method to create synthetic data for medical machine learning.
• Synthetic data augmentation enhances model robustness and fairness, crucial for real-world clinical deployment.
• This approach shows promise in mitigating performance gaps caused by data underrepresentation in healthcare AI.