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

This paper explores the use of adversarial debiasing algorithms to mitigate bias in ultrasound imaging datasets, focusing on breast and lung images. The study evaluates fairness metrics like area under the curve (AUC), False Positive Rate (FPR), False Negative Rate (FNR), and demographic parity to assess the impact of debiasing using the recently published MEDFAIR framework. While debiasing improves fairness overall, disparities remain in certain subgroups, such as age in the breast dataset and sex in the lung dataset. The paper also compares artificial intelligence (AI) models (ResNet18, AlexNet, VGG16, MobileNetV2, DenseNet121), revealing differences in susceptibility to bias and effectiveness post-debiasing. These findings underscore the challenges of achieving full fairness in AI-driven medical imaging and highlight the need for continued refinement of debiasing methods to ensure equitable outcomes across diverse patient populations.