Robust Statistical Inference in Human Brain Mapping

Abstract

Statistical parametric mapping has been widely used in human brain mapping to explain brain image changes as a function of other factors. The core theory underlying this approach is the general linear model (GLM). Originally introduced for structural magnetic resonance image (MRI) and positron emission tomography analysis, this framework has been extended to resting state functional MRI and multi-modality brain image analysis. Despite the power of the extensions, problems within the traditional GLM assumptions and ordinary least squares (OLS) estimation arise.

The aim of this dissertation is to develop robust and accurate models within the GLM framework for multi-modality brain mapping and functional connectivity analysis. We introduced and modified modern statistical methods, which are established in statistical community, in the context of human brain mapping to obtain robust and accurate estimations. The robust regression and non-parametric mapping were introduced to address outlier problems. Model II regression and regression calibration were introduced to consider the imaging regressors in multi-modality brain image analysis. We developed spatial temporal models to account for spatial and temporal correlations simultaneously for functional connectivity analysis. To evaluate our methods, we proposed a quantitative approach for comparing inference methods on empirical studies. A large multi-site study was conducted to investigate the application of inter-modality human brain mapping using a shared database.