DeepResBat: Deep residual batch harmonization accounting for covariate distribution differences

Lijun An, Chen Zhang, Naren Wulan, Shaoshi Zhang, Pansheng Chen, Fang Ji, Kwun Kei Ng, Christopher Chen, Juan Helen Zhou, B. T. Thomas Yeo

Abstract: Highlights•Theoretical analysis shows that ignoring covariates may lead to poor harmonization.•Deep learning may hallucinate relationships between covariates and MRI measurements.•Proposed residual harmonization framework DeepResBat mitigates false positive issue.•DeepResBat outperforms classical mixed-effect and deep learning approaches.•Exhaustive evaluation on three large datasets with 10,085 anatomical T1 scans.