Fig. 1: A comprehensive framework of our study.
From: Enhancing fairness in AI-enabled medical systems with the attribute neutral framework
a Utilizing the AttrNzr to create modified X-ray images, where the attributes of these modified images can be regulated by the modification intensity α. b Involving AI judges and human judges in the attribute recognition to discriminate the original attributes of the modified X-ray images. c Establish a DDM on neutralized X-ray images generated by the AttrNzr with an α value of 0.5. d Conduct performance and unfairness evaluations based on predictions from the DDM. e Investigate the implementation of AttrNzrs in both the test-stage neutralization paradigm to mitigate unfairness in pre-existing medical AI models, as well as the training-stage neutralization paradigm to minimize computational requirements in new medical AI models.
