Model Based Principal Component Analysis with Application to Functional Magnetic Resonance Imaging.

Abstract

Functional Magnetic Resonance Imaging (fMRI) has allowed better understanding of human brain organization and function by making it possible to record either autonomous or stimulus induced brain activity. After appropriate preprocessing fMRI produces a large spatio-temporal data set, which requires sophisticated signal processing. The aim of the signal processing is usually to produce spatial maps of statistics that capture the effects of interest, e.g., brain activation, time delay between stimulation and activation, or connectivity between brain regions. Two broad signal processing approaches have been pursued; univoxel methods and multivoxel methods. This proposal will focus on multivoxel methods and review Principal Component Analysis (PCA), and other closely related methods, and describe their advantages and disadvantages in fMRI research. These existing multivoxel methods have in common that they are exploratory, i.e., they are not based on a statistical model. A crucial observation which is central to this thesis, is that there is in fact an underlying model behind PCA, which we call noisy PCA (nPCA). In the main part of this thesis, we use nPCA to develop methods that solve three important problems in fMRI. 1) We introduce a novel nPCA based spatio-temporal model that combines the standard univoxel regression model with nPCA and automatically recognizes the temporal smoothness of the fMRI data. Furthermore, unlike standard univoxel methods, it can handle non-stationary noise. 2) We introduce a novel sparse variable PCA (svPCA) method that automatically excludes whole voxel timeseries, and yields sparse eigenimages. This is achieved by a novel nonlinear penalized likelihood function which is optimized. An iterative estimation algorithm is proposed that makes use of geodesic descent methods. 3) We introduce a novel method based on Stein’s Unbiased Risk Estimator (SURE) and Random Matrix Theory (RMT) to select the number of principal components for the increasingly important case where the number of observations is of similar order as the number of variables.