Independent Component versus Local Sparse Component Analysis in Resting State fMRI

Abstract

Independent Component Analysis (ICA) algorithms are potentially powerful ways of localizing sources of cerebral activity in resting state functional Magnetic Resonance Imaging (fMRI). But the assumptions underling the nature of identified sources limits this tool. By creating local one-dimensional approximations, Local Sparse Component Analysis (LSCA) can separate contiguous sources on the basis of their sparse representation into smoothness spaces via the 3D wavelet transformation. In this paper we systematically compare Probabilistic ICA (PICA) and LSCA for analyzing resting state fMRI across healthy participants. We show that the PICA sources usually representing biologically plausible components can in fact be decomposed into several LSCA sources that are not necessarily independent from each other. In addition, we show that LSCA identifies sources that approximate much better the local variations of the blood oxygenation level-dependent (BOLD) signal than PICA sources.