'Harmless' adversarial network harmonization approach for removing site effects and improving reproducibility in neuroimaging studies.

Abstract

Multi-site collaboration, which gathers together samples from multiple sites, is a powerful way to overcome the small-sample problem in the neuroimaging field and has the potential to discover more robust and reproducible biomarkers. However, confounds among the datasets caused by various site-specific factors may dramatically reduce the cross-site reproducibility performance. To properly remove confounds while improving cross-site task performances, we propose a maximum classifier discrepancy generative adversarial network (MCD-GAN) that combines the advantages of generative models and maximum discrepancy theory. The mechanisms of MCD-GAN and how it harmonizes the dataset are visualized using simulated data. The performance of MCD-GAN was also compared with state-of-the-art methods (e.g., ComBat, cycle-GAN) within Adolescent Brain Cognitive Development (ABCD) dataset. Result demonstrates that the proposed MCD-GAN can effectively improve the cross-site gender classification performance by harmonizing site effects. Our proposed framework is also suitable for various classification/prediction tasks and is promising to facilitate the cross-site reproducibility of neuroimaging studies. Clinical Relevance- This work provides an efficient method for removing sites effects and improving reproducibility in large-cohort neuroimaging studies.