AI alignment in medical imaging: Unveiling hidden biases through counterfactual analysis

Area of Science:

• Medical Imaging AI
• Machine Learning Fairness
• Statistical Modeling

Background:

• Machine learning (ML) models in medical imaging show high diagnostic accuracy.
• These models risk learning spurious correlations with sensitive attributes (e.g., demographics), impacting fairness and safety.
• Evaluating and mitigating such biases is crucial for trustworthy AI in healthcare.

Purpose of the Study:

• To develop a novel statistical framework for evaluating the dependency of medical imaging ML models on sensitive attributes.
• To quantify the extent to which model predictions are invariant to hypothetical changes in sensitive attributes.
• To provide a practical method for identifying and measuring bias in AI diagnostic tools.

Main Methods:

• Leveraged the concept of counterfactual invariance to measure bias.
• Developed a practical algorithm combining conditional latent diffusion models with statistical hypothesis testing.
• Evaluated the framework on synthetic benchmarks and real-world chest X-ray datasets (CheXpert, MIMIC-CXR).

Main Results:

• On synthetic data, the framework achieved 96.1% accuracy in identifying biased models with a 22.0% false alarm rate.
• On real-world datasets, bias detection rates reached 96.3% and 95.7% for strongly biased models.
• False alarm rates for less biased models were 15.3% and 14.7%, outperforming existing fairness baselines.

Conclusions:

• The proposed framework effectively identifies and quantifies bias in medical imaging ML models.
• It demonstrates strong alignment with counterfactual fairness principles.
• The method offers a reliable approach to enhance fairness and safety in clinical AI applications.