Structured sparse principal component analysis for fMRImaging

The use of component analysis on fMRI data allows to extract some interesting hidden components out of the data such as neuronal networks. While independent component analysis (ICA) is currently preferred in this application, we try to know whether sPCA could give better results than ICA when applied to fMRI data by highlighting the strengths and weaknesses of both techniques. Indeed, the neuronal networks are mostly intrinsically very sparse and it could thus be interesting to include explicitely this feature in the decomposition technique. In the experimental section we first show on simulated fMRI data that in an ideal example of fMRI data, sPCA gives better results than ICA when the sparsity of the networks composing the simulated data is higher than approximately 80%. However, using the same model it appears that sPCA seems to be less robust than ICA to some perturbations that we can find in real fMRI data such as the motion of the patient during acquisition of the data. We then use real fMRI data and we design three different experiments in order to evaluate the decomposition performed by both techniques. In each experiment ICA gives better results than sPCA. We can retain two important drawbacks of sPCA compared to ICA. First, the neuronal networks extracted through sPCA appear to be more affected by perturbations such as the motion of the patient, making the extraction of neuronal components from one subject to another less robust than with ICA. Second, sPCA does not seem to be able to isolate neuronal information in a few components only, whereas ICA does.