B1049
Title: Statistical harmonization for the neuroimaging data with complex data distributions
Authors: Andrew Chen - University of Pennsylvania (United States)
Haochang Shou - University of Pennsylvania (United States) [presenting]
Russell Shinohara - University of Pennsylvania (United States)
Abstract: With the increasing need for big data analytics in medical imaging, pooling and integrating data from multi-site studies has become critical. Site differences attributed to various sources are known to exist and might result in a substantial impact on the analytic results. Recently, batch-effect correction methods such as ComBat and CovBat have been successfully adapted to remove scanner and site differences in multimodal neuroimaging data and applied in many large-scale studies. However, the model assumptions of the existing statistical harmonization methods might restrict their direct application to broader imaging modalities with complex distributions. For example, fewer methods are available to harmonize the resting-state functional magnetic resonance imaging (fMRI) connectivity matrices, given the complex dependency structures temporally and spatially in the raw fMRI data and that the derived connectivity matrices do not necessarily belong to Euclidean metric space. Additionally, the current ComBat or CovBat assumes a Gaussian residual error and does not apply to imaging measures with skewed distribution or zero abundance such as white matter lesion counts. We will discuss several extensions of the statistical harmonization methods to complex neuroimaging modalities including multisite functional connectivity data and white matter hyperintensity data. The methods are shown to promote a more robust community detection in network analysis.