BARTharm: MRI Harmonization Using Image Quality Metrics and Bayesian Non-parametric

Abstract

Image derived phenotypes (IDPs) harmonization from Magnetic Resonance Imaging (MRI) data is essential for reducing scanner-induced, non-biological variability and enabling accurate multi-site analysis. Existing methods like ComBat, while widely used, rely on linear assumptions and explicit scanner IDs - limitations that reduce their effectiveness in real-world scenarios involving complex scanner effects, non-linear biological variation, or anonymized data. We introduce BARTharm, a novel harmonization framework that uses Image Quality Metrics (IQMs) instead of Scanner IDs and models scanner and biological effects separately using Bayesian Additive Regression Trees (BART), allowing for flexible, data-driven adjustment of IDPs. Through extensive simulation studies, we demonstrate that IQMs provide a more informative and flexible representation of scanner-related variation than categorical Scanner IDs, enabling more accurate removal of non-biological effects. Leveraging this and its ability to model complex relationships, BARTharm, consistently outperforms ComBat across a range of challenging scenarios, including model misspecification and confounded scanner-biological relationships. Applied to real-world datasets, BARTharm successfully removes scanner-induced bias while preserving meaningful biological signals, resulting in stronger, more reliable associations with clinical outcomes. Overall, we find that BARTharm is a robust, data-driven improvement over traditional harmonization approaches, particularly suited for modern, large-scale neuroimaging studies.