A supervised singular value decomposition for independent component analysis of fMRI

Abstract

Functional Magnetic Resonance Imaging (fMRI) is a non-invasive technique for studying the brain activity. The data acquisition process results a temporal sequence of 3D brain images. Due to the high sensitivity of MR scanners, spikes are commonly observed in the data. Along with the temporal and spatial features of fMRI data, this artifact raises a challenging problem in the statistical analysis. In this paper, we introduce a supervised singular value decomposition technique as a data reduction step of independent component analysis (ICA), which is an effective tool for exploring spatio-temporal features in fMRI data. Two major advantages are discussed: first, the proposed method improves the robustness of ICA against spikes; second, the method uses the fMRI experimental designs to guide the fully data-driven ICA, yielding a more computationally efficient procedure and highly interpretable results. The advantages are demonstrated using spatio-temporal simulation studies as well as a data analysis.