Model-free Region Of Interest Based Analysis of fMRI Data I.R. Keck, F.J. Theis, P. Gruber, E.W. Lang Institute of Biophysics University of Regensburg 93040 Regensburg, Germany elmar.lang@biologie.uni-regensburg.de J. Churan Generation Research Programm LMU Munich 83646 Bad T¨ olz, Germany churan@grp.hwz.uni-muenchen.de C. G. Puntonet Departamento ATC Universidad de Granada/ESII E-1807 Granada, Spain carlos@atc.ugr.es Abstract— Blind source separation is beeing established as a new kind of analysis method in the field of fMRI data analysis. Its model-free approach renders it an important tool for investigation. In this article we present a method based on BSS/ICA to create a region of interest (ROI) mask for a BSS based fMRI data analysis. This eliminates a commonly necessary model-based step in fMRI data analysis. We demonstrate our approach on a real world fMRI data set example. I. I NTRODUCTION Functional magnetic resonance imaging (fMRI) is one of the fundamental tools for functional human brain research due to its high spatial resolution and noninvasiveness. Nowadays Blind Source Separation (BSS) methods like Independent Component Analysis (ICA) become standard methods of analysing functional magnetic resonance data (see e.g. [1] for a recent review). However, a problem in case of fMRI data still is the amount of data that has to be analyzed to find interesting components: Each fMRI session will typically yield hundreds of different components. The regions of the brain that work together during the experiment will form clusters of activation in the time do- main. These clusters can be exploited by grouping together those components which represent the collaborating parts of the brain. Two recently published algorithms to tackle these problem are tree-dependent ICA [2] and topographic ICA [3]. In tree-dependent ICA the assumption of strict independence is relaxed by transforming the data into a tree of independent clusters of dependent sources. In topographic ICA also the re- sulting components do not have to be completely independent: the variances corresponding to neighboring components have to be positively correlated while the other variances remain independent. Both algorithms have been applied to fMRI data by [4] with varying degrees of success. However, in the search for cooperating networks of active brain areas in fMRI, underlying independent components need not to be similar themselves. Rather the time-courses of their activations, i.e. the corresponding columns of the mixing matrix estimated with ICA, will show similarities. In this article we therefore follow an alternative route: • First we use a temporal SOBI algorithm as described in [5] to identify interesting networks in the fMRI data. fMRI data set tICA for ROI mask-generation masking search for networks activation patterns sICA Fig. 1. The workflow of an fMRI analysis using temporal ICA to generate ROIs for a subsequent spatial ICA. • Second we utilise these networks as region-of-interest (ROI) masks to filter out the interesting parts of the data set and use the FastICA algorithm [6] to apply spatial ICA on these ROIs. II. THEORY First we present a short overview on independent component analysis. Then we will describe how to utilise a temporal ICA to generate a ROI mask for the subsequent spatial ICA. A. Independent Component Analysis Let s 1 (t),...,s m (t) designate m independent signals with unit variance for simplicity, represented by a vector s(t)= (s 1 (t),...,s m (t)) T , where T denotes the transpose. Let the mixing matrix A generate n linear mixtures x(t) = (x 1 (t),...,x n (t)) T from these source signals according to: x(t)= As(t)= m  i=1 s i (t) a i (1)