Enhancing Fairness in Ultrasound Imaging: Evaluating Adversarial Debiasing Across Diverse Patient Demographics

Area of Science:

• Medical Imaging AI
• Machine Learning Fairness
• Ultrasound Technology

Background:

• AI models in medical imaging can perpetuate societal biases present in training data.
• Ensuring fairness in AI diagnostic tools is critical for equitable patient care.
• Ultrasound datasets, particularly for breast and lung imaging, may contain demographic disparities.

Purpose of the Study:

• To evaluate the effectiveness of adversarial debiasing algorithms in mitigating bias within breast and lung ultrasound datasets.
• To assess the impact of debiasing on fairness metrics across different AI models.
• To identify remaining fairness disparities in AI models after debiasing.

Main Methods:

• Utilized adversarial debiasing techniques within the MEDFAIR framework.
• Evaluated fairness using metrics: Area Under the Curve (AUC), False Positive Rate (FPR), False Negative Rate (FNR), and demographic parity.
• Compared bias and debiasing performance across various AI architectures (ResNet18, AlexNet, VGG16, MobileNetV2, DenseNet121).

Main Results:

• Adversarial debiasing improved overall fairness in ultrasound AI models.
• Residual disparities were observed in specific subgroups: age (breast) and sex (lung).
• Different AI models exhibited varying susceptibility to bias and effectiveness post-debiasing.

Conclusions:

• Achieving complete fairness in AI-driven medical imaging remains a significant challenge.
• Current debiasing methods show promise but require further development to address subgroup disparities.
• Ongoing research into advanced debiasing strategies is essential for equitable AI in healthcare.