Online Dictionary Learning For Single-Subject Fmri Data Unmixing

Independent component analysis (ICA) and dictionary learning (DL) methods are widely used to analyse resting state functional Magnetic Resonance Imaging (rs-fMRI) in multi-subject studies. These methods aim at decomposing the multi-subject data into common spatial abundance maps and their related temporal signatures.We are interested here in such a decomposition for a single-subject rs-fMRI dataset. The above-mentioned methods often fail in this case because the problem becomes too ill-posed, requiring the use of additional prior information and the design of novel regularising constraints. The poor resolution of rs-fMRI data is an additional source of difficulty, yielding noisy and blurry spatial maps.In this paper, we propose a new DL formulation adapted to the unique subject by integrating high-resolution (HR) spatial information to constrain single-subject data unmixing. HR information is provided by the registration of an anatomical atlas on the data set. We show on a quasi-real dataset from mice, the benefit of using an HR spatial segmentation map in the decomposition of low-resolution rs-fMRI.