FairREAD: Re-fusing demographic attributes after disentanglement for fair medical image classification

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Deep learning shows promise in medical imaging but suffers from performance disparities across demographic subgroups.
• Current bias mitigation techniques often remove sensitive attributes, potentially compromising clinically relevant information and overall model performance.

Purpose of the Study:

• To introduce Fair Re-fusion After Disentanglement (FairREAD), a novel framework designed to mitigate unfairness in deep learning models for medical imaging.
• To address the challenge of balancing bias reduction with the preservation of diagnostic accuracy.

Main Methods:

• FairREAD disentangles demographic information using orthogonality constraints and adversarial training.
• A controlled re-fusion mechanism reintegrates sensitive attributes into fair image representations, preserving clinical relevance.
• Subgroup-specific threshold adjustments are employed to ensure equitable performance across different demographic groups.

Main Results:

• Comprehensive evaluations on large-scale clinical X-ray datasets demonstrated significant reduction in unfairness metrics.
• FairREAD maintained high diagnostic accuracy across demographic subgroups.
• Out-of-distribution testing confirmed the robustness of the proposed framework.

Conclusions:

• FairREAD offers an effective solution for mitigating demographic bias in medical imaging AI.
• The framework successfully balances fairness and diagnostic performance, addressing a critical limitation in current deep learning applications.
• The approach preserves clinically relevant information while ensuring equitable outcomes for diverse patient populations.