Racial Disparities Persist Beyond Data Representation in Medical Imaging — even Predictive Uncertainty Fails to Capture them

Tareen Dawood, Gloria Stucchi, Aasa Feragen

Keywords: Algorithmic Fairness, Bias, Uncertainty, Disparities, Representation

TL;DR: While racial training set representation affects model performance, there is more at play, as large racial disparities remain regardless of training set composition.

Abstract:

Balanced training sets are often promoted to mitigate racial performance disparities of
Deep Learning (DL) models in medical imaging. However, our preliminary findings on
two medical imaging datasets show that while racial training set representation
affects model performance, there is more at play, as large racial disparities
remain regardless of training set composition. Moreover, predictive uncertainty is
shown to be completely insensitive to these performance disparities. From this, we derive
a series of open problems for safe and fair image-guided diagnostics.

Submission Number: 46