A0448
Title: Promises of covariance harmonization in multi-site neuroimaging studies
Authors: Jun Young Park - University of Toronto (Canada) [presenting]
Abstract: It is well-known that batch effects severely reduce data quality in modern neuroimaging or genomic studies. In response, many statistical methods have been developed in the past decade to homogenize batch effects for effective downstream statistical analyses, exemplified by the ComBat method that models heterogeneity in means and variances. It is shown that modeling heterogeneity in covariances (in addition to means and variances) substantially improves the quality of batch effect correction in neuroimaging studies. At the same time, there are multiple approaches to model covariance heterogeneity, which motivates a solid understanding of data characteristics for a more successful harmonization. RELIEF and SAN are showcased, which use a low-rank factor model or a spatial Gaussian process to model covariance batch effects in various neuroimaging data types, such as diffusion tensor imaging (DTI) or cortical thickness. The empirical performance of these methods are demonstrated using the SPINS study and discuss a few possible extensions of these methods.
