Adversarial Debiasing With Partial Learning - Medical Image Studies

Explore adversarial debiasing techniques for reducing racial disparities in medical image AI models through this 33-minute conference talk by Ramon Correa from Stanford University. Delve into the challenges of implementing trustworthy clinical AI models and the issue of implicit biases in decision-making processes. Learn about a novel two-step adversarial debiasing approach with partial learning, designed to mitigate racial disparity while maintaining model performance. Examine case studies on chest X-rays and mammograms that demonstrate the potential of this methodology. Gain insights into the speaker's research on model debiasing techniques and the importance of addressing biases in healthcare AI applications. Understand the background of bias in machine learning models, the impact of racial disparities in medical AI, and potential solutions such as balanced dataset training and adversarial debiasing architectures. Discover the findings on predicting race from medical images and the effectiveness of removing race-related features. Analyze the results of debiasing efforts on chest X-ray models and their implications for future research in the field of medical AI.

Intro
Bias in ML models
Background: Studying Biases in Medical AI models
These Algorithms Look at X-Rays-and Somehow Detect Your Race
Solution: Retrain Model With Balance Dataset
Emory CXR dataset Balanced Training
TPR disparities persist in "balanced" datasets
Adversarial debiasing: Unlearn Biasing features
Adversarial Debiasing Architecture
Adversarial Debiasing Background
Ablation Studies: Review
How do we identify the layers to debias?
Emory Mammogram Dataset • Cohorts of 150-180k patients each featuring screening
Emory Mammography Dataset (Race Distribution )
Deep learning model for tissue density classification
Studying Models Ability to predict Race
Findings on Predicting Race
Removing Race Related Features (Step 2)
TPR Disparity Measures
CXR Model Performance
Questions?
CXR Debiasing Results