[Paper Review] Two-step adversarial debiasing with partial learning -- medical image case-studies

Ramón Correa, Jiwoong Jeong+4 more|arXiv (Cornell University)|Nov 16, 2021

Adversarial Robustness in Machine Learning4 citations

TL;DR

This paper proposes a two-step adversarial debiasing method with partial fine-tuning to reduce racial bias in medical image AI models while preserving diagnostic performance. It achieves significant disparity reduction—e.g., TPR disparity for edema in chest X-rays dropped from 1.33 to 1.11—without sacrificing task accuracy, outperforming full debiasing in key cases like mammography for Asian patients.

ABSTRACT

The use of artificial intelligence (AI) in healthcare has become a very active research area in the last few years. While significant progress has been made in image classification tasks, only a few AI methods are actually being deployed in hospitals. A major hurdle in actively using clinical AI models currently is the trustworthiness of these models. More often than not, these complex models are black boxes in which promising results are generated. However, when scrutinized, these models begin to reveal implicit biases during the decision making, such as detecting race and having bias towards ethnic groups and subpopulations. In our ongoing study, we develop a two-step adversarial debiasing approach with partial learning that can reduce the racial disparity while preserving the performance of the targeted task. The methodology has been evaluated on two independent medical image case-studies - chest X-ray and mammograms, and showed promises in bias reduction while preserving the targeted performance.

Motivation & Objective

• To address demographic bias in clinical AI models, particularly racial disparities in medical image classification.
• To develop a debiasing method that maintains high performance on target diagnostic tasks despite reducing reliance on protected attributes like race.
• To evaluate whether partial fine-tuning after adversarial debiasing can further improve performance while reducing bias.
• To compare full vs. partial debiasing techniques across two real-world medical imaging use cases: chest X-ray and mammography.
• To provide a scalable, architecture-agnostic approach to debiasing that works with gradient-based training.

Proposed method

• The method uses a dual-branch CNN architecture: one branch predicts the target diagnostic label (y), and the other adversarially predicts the protected attribute (Z, e.g., race).
• During training, the model minimizes a combined loss: L = L_predictor + λ × L_adversarial, with λ = 0.53, to jointly learn both tasks.
• In a second backward pass, the gradient from the adversarial branch is flipped (L = -λ × L_adversarial) to penalize learning of protected attributes.
• Partial fine-tuning is applied by identifying and retraining top 10% of filters in convolutional layers that most degrade protected attribute prediction, using ablation analysis.
• The approach is applied post-adversarial training, allowing performance recovery while maintaining bias reduction.
• The method is compatible with any differentiable model architecture and does not require demographically balanced datasets.

Experimental results

Research questions

• RQ1 Can adversarial debiasing reduce racial disparities in medical image classification without degrading diagnostic performance?
• RQ2 Does partial fine-tuning after adversarial debiasing improve model performance while maintaining fairness?
• RQ3 How do full vs. partial debiasing strategies compare in terms of bias reduction and task accuracy across different imaging modalities?
• RQ4 To what extent does the model’s reliance on race correlate with diagnostic outcomes in chest X-ray and mammogram interpretation?
• RQ5 Can this method be generalized across different medical imaging tasks without requiring retraining from scratch?

Key findings

• For chest X-ray, partial debiasing improved AUC for no finding from 0.866 to 0.889 and reduced TPR disparity for edema from 1.33 to 1.11, bringing it within the acceptable 0.8–1.25 range.
• In mammography, partial debiasing reduced TPR disparity for Asian patients from 3.37 to 1.67 on low-density tissue classes, significantly improving fairness.
• The partial debiasing method preserved or slightly improved performance on the target task, while full debiasing degraded performance in some cases (e.g., AUC dropped from 0.965 to 0.942 for tissue density class 1).
• For heterogeneously dense tissue in mammography, no significant disparity improvement was observed for Asian patients, likely due to high prevalence of dense breasts in this group.
• The method successfully reduced bias in both datasets while maintaining or improving diagnostic performance, particularly in cases where baseline models showed high disparity.
• The ablation-based partial fine-tuning strategy identified key convolutional layers that, when retrained, restored performance lost during adversarial debiasing.