Localisation of Racial Information in Chest X-Ray for Deep Learning Diagnosis

Deep learning-based classification of diseases from Chest X-ray has been shown to use implicit information about the subjects’ self-reported race, which could result in diagnostic bias. In this paper, we describe and compare two approaches to investigate where racial information is located in the image: first leveraging non-linear registration and computing atlas differences and second using saliency maps. We compute a map visualising the racial information between black and white subjects and discuss whether those maps are consistent with the model explanation.