Over the past decade, the number of large multi-center neuroimaging studies has skyrocketed due to growing investments by federal governments and private entities interested in brain development, aging, and pathology. This has led to the accumulation of vast amounts of magnetic resonance imaging (MRI) data which have been acquired with varying amounts of technical harmonization. Such efforts, which have focused on protocol harmonization and comparisons with imaging phantoms, have shown great strides toward reducing inter- scanner differences in imaging features extracted for further study. Unfortunately, MRI show inter-instrument biases even in the most carefully controlled studies. Our group, among many others, has shown that these differences often dwarf biological differences of interest measured using both structural and functional MRI. To address this, the field has rapidly been developing tools for the harmonization of imaging data after acquisition. We have proposed several such tools, and our work has often focused on the adaptation of methods used in genomic studies for batch effect correction. Our most recent such work involved the ComBat method, which uses empirical Bayesian estimation to correct for site effects in both means and variances of imaging features under study. To date, these tools have been successfully applied in studies of cortical thickness, white matter microstructure, and functional connectivity. However, there are unfortunately several key limitations to the ComBat method for imaging studies that stem from its original conception for gene expression studies. ComBat was designed for the study of inter-scanner differences in cross-sectionally acquired data. While cross-sectional studies are of great interest and exceedingly common, much focus in the context of healthy brain development and aging has shifted to measuring longitudinal trajectories. In such cases, the nave application of ComBat is flawed and methodological research is necessary for appropriate harmonization tools to be developed. Furthermore, more complex nested study design in which multiple scanners are used per institution, or a subset of subjects are imaged on multiple scanners for harmonization purposes, are increasingly common. Another key area of interest in modern neuroimaging studies is to focus on inter-region structural or functional connectivity and uses multivariate pattern analysis (MVPA) to improve our understanding of phenotypic associations as well as for personalized predictions. Unfortunately, the current state-of-the-art in image harmonization ignores correlation structure between measurements, and thus inter- scanner differences often persist. In this project, we propose a new generation of techniques that are applicable under complex study designs and harmonize appropriately for studies involving applications of MVPA. In our final aim of this proposal, we will apply the methods developed for more complex study designs and MVPA in the context of two of the largest NIH-funded multi-center consortia across the lifespan.
Over the past decade, the number of large multi-center neuroimaging studies has skyrocketed due to growing investments by federal governments and private entities interested in brain development, aging, and pathology. This has led to the accumulation of vast amounts of magnetic resonance imaging (MRI) data acquired using various study designs, but the statistical methodology necessary for harmonizing these data for integrated analyses is currently lacking. In this project, we propose to develop, implement, and apply next-generation imaging harmonization methods for data acquired longitudinally or from more complex study designs, as well as appropriate harmonization methods for use in conjunction with popular multivariate pattern analysis techniques.
