Improving Fairness in Chest X-Ray Interpretation Models Using Attention-Driven Masked Image Modeling.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Health Equity

Background:

• Growing concerns exist regarding demographic biases in artificial intelligence (AI) models used in healthcare.
• AI models may perpetuate or even amplify existing health disparities.
• Ensuring equitable AI performance across diverse patient populations is critical for medical applications.

Purpose of the Study:

• To investigate demographic bias in Vision Transformer (ViT) models for chest X-ray (CXR) analysis.
• To develop and evaluate a novel method, Attention-driven Masked Image Modeling (AMIM), to mitigate identified biases.
• To enhance the fairness and equity of AI in medical diagnostic tools.

Main Methods:

• Utilized a Vision Transformer (ViT) model for multi-label classification of CXRs.
• Employed transfer learning to assess the model's ability to discern racial demographics.
• Analyzed model attention heatmaps to identify regions potentially encoding demographic information.
• Developed and applied Attention-driven Masked Image Modeling (AMIM) with masked patches to reduce bias.

Main Results:

• The ViT model achieved an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.72 for CXR classification.
• The model demonstrated an AUC of 0.66 in discerning racial demographics, despite no demographic data in training.
• Attention heatmap analysis highlighted image regions potentially linked to demographic information.
• AMIM approach with 9 masked patches resulted in an AUC of 0.71, indicating reduced demographic disparities.

Conclusions:

• Vision Transformer models can inadvertently learn demographic information from medical images.
• Attention-driven Masked Image Modeling (AMIM) shows promise in mitigating demographic biases in AI-driven medical imaging.
• This research contributes to developing more equitable and reliable AI tools for healthcare applications.