Efficient adversarial debiasing with concept activation vector - Medical image case-studies.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning Bias

Background:

• AI models face challenges in real-time deployment due to trustworthiness issues with unseen populations.
• Complex AI models often act as black boxes, exhibiting implicit biases in decision-making, especially for minority subgroups.

Purpose of the Study:

• To develop an efficient adversarial de-biasing approach for AI models.
• To reduce racial disparities in AI decision-making while preserving task performance.
• To adapt model interpretability techniques for bias mitigation.

Main Methods:

• Developed an adversarial de-biasing approach using concept activation vectors (CAV).
• Utilized CAV to identify convolution layers responsible for learning race.
• Applied partial learning by fine-tuning only identified layers, minimizing performance drop.

Main Results:

• Evaluated on chest X-ray and mammogram datasets with external validation.
• Debiased chest X-ray models achieved higher AUC (0.87) than baseline (0.57) and fine-tuned models (0.81).
• Mammogram models improved performance on external datasets (AUC 0.8 to 0.86) after debiasing.

Conclusions:

• Adversarial models trained internally performed comparably or better than standard fine-tuning with external data.
• The adversarial training approach is model-architecture agnostic for gradient-based models.
• Training code is available under an academic open-source license.