摘要
Ensuring both accuracy and fairness in AI-assisted chest X-ray (CXR) diagnostics is critical, particularly given the presence of sensitive patient information. In this study, we introduce ShortCXR, a Shortcut-Infused Chest X-Ray Dataset explicitly designed to expose and measure biases stemming from intrinsic image features, data sources, and demographic factors. We then integrate self-supervised learning (SSL) techniques into a series of experiments to evaluate how these biases affect diagnostic performance. Our findings demonstrate that state-of-the-art SSL methods boost both accuracy and fairness, effectively mitigating the impact of distortion-based, source-based, and demographic biases. By highlighting SSL's potential to enhance diagnostic quality and equity, this work provides a strong foundation for future research on bias mitigation in medical imaging.