Mitigating Racial Bias in Chest X-Ray Disease Diagnosis Via Disentanglement Learning

The application of artificial intelligence (AI) has achieved remarkable advancements in medical imaging. However, due to unbalanced distribution, most existing diagnoses suffer from racial bias, where diagnosis model tends to achieve different performance among races. This racial bias in performance seriously affects the credibility of diagnosis model, and eventually hinders the application of new technologies. In this work, to address the issue of unfairness, we propose a novel debias model to disentangle and utilize disease features that are apart from racial features for disease diagnosis. Results show that the proposed model could reduce the unfairness caused by racial factors in imbalanced data, and significantly mitigate racial bias. Source codes are available at Github¹.