PhyCHarm : Physics-Constrained Deep Neural Networks for Multi-Scanner Harmonization

Motivation: The MR scanner effect in a multi-site dataset can affect bias in statistical analysis or reduce generality in deep neural networks. Goal(s): We aim to suggest a MR physics-informed harmonization framework (PhyCHarm) that generates consistent quantitative maps and harmonized T1w images. Approach: We introduce a Quantitative Maps Generator and a Harmonization Network to be trained with a constraint loss based on a signal equation. Results: PhyCHarm shows the highest evaluation scores in both networks and consistent segmentation accuracy in the downstream task (FSL FAST GM and WM segmentation). Impact: PhyCHarm works based on the Bloch equation. PhyCHarm enables us to reduce scanner effects efficiently in the dataset before conducting test/retest, longitudinal, or multi-site studies. It can be helpful to ensure deep neural networks' generality.