Large-scale optimization for component analysis of fMRI resting brain data

The use of component analysis on fMRI data is an important neuroimaging computational tool. In this paper we focus on the particular application of extracting the so-called default mode neuronal network from resting brain data ([1] and references therein). While independent component analysis (ICA) is currently the method of choice in this application, we investigate the advantages and limitations of using sparse PCA as an alternative to ICA. Indeed, the searched neuronal networks are mostly intrinsically very sparse and it has been suggested that ICA is a prior for sparsity rather than for statistical independence in neuroimaging [2].