Comparing EEG/MEG neuroimaging methods based on localization error, false positive activity, and false positive connectivity

Abstract

1.AbstractEEG/MEG neuroimaging consists of estimating the cortical distribution of time varying signals of electric neuronal activity, for the study of functional localization and connectivity. Currently, many different imaging methods are being used, with very different capabilities of correct localization of activity and of correct localization of connectivity. The aim here is to provide a guideline for choosing the best (i.e. least bad) imaging method. This first study is limited to the comparison of the following methods for EEG signals: sLORETA and eLORETA (standardized and exact low resolution electromagnetic tomography), MNE (minimum norm estimate), dSPM (dynamic statistical parametric mapping), and LCMVBs (linearly constrained minimum variance beamformers). These methods are linear, except for the LCMVBs that make use of the quadratic EEG covariances. To achieve a fair comparison, it is assumed here that the generators are independent and widely distributed (i.e. not few in number), giving a well-defined theoretical population EEG covariance matrix for use with the LCMVBs. Measures of localization error, false positive activity, and false positive connectivity are defined and computed under ideal no-noise conditions. It is empirically shown with extensive simulations that: (1) MNE, dSPM, and all LCMVBs are in general incapable of correct localization, while sLORETA and eLORETA have exact (zero-error) localization; (2) the brain volume with false positive activity is significantly larger for MN, dSPM, and all LCMVBs, as compared to sLORETA and eLORETA; and (3) the number of false positive connections is significantly larger for MN, dSPM, all LCMVBs, and sLORETA, as compared to eLORETA. Non-vague and fully detailed equations are given. PASCAL program codes and data files are available. It is noted that the results reported here do not apply to the LCMVBs based on EEG covariance matrices generated from extremely few generators, such as only one or two independent point sources.