ICA-based artifact correction improves spatial localization of adaptive spatial filters in MEG

Beamformers are one of the most common inverse models currently used in the estimation of source activity from magnetoencephelography (MEG) data. They rely on a minimization of total power while constraining the gain in the voxel of interest, resulting in the suppression of background noise. Nonetheless, in cases where background noise is strong compared to the source of interest, or when many sources are present, the ability of the beamformer to detect and accurately localize weak sources is reduced. In visual paradigms, two main background sources can substantially impact an accurate estimation of weaker sources. Ocular artifacts are orders of magnitude higher than neural sources making it difficult for the beamformer to effectively suppress them. Primary visual activations also result in strong signals that can impede localization of weak sources. In this paper, we systematically evaluated how neural (visual) and non-neural (eye, heart) sources affect the localization accuracy of frontal and medial temporal sources in visual tasks. These sources are of tremendous interest in learning and memory studies as well as in clinical settings (Alzheimer's/epilepsy) and are typically difficult to localize robustly in MEG. Empirical data from two tasks - active learning and control - were used to evaluate our analysis techniques. Global field power calculations showed multiple time periods where active learning was significantly different from response selection with dominant sources converging to the eyes. Extensive leakage of eye activity into frontal and visual that evoked responses into parietal cortices was also observed. Contributions from ocular activity to the reconstructed time series were indiscernible from task-based recruitment of frontal sources in the original data. Removing artifacts (eye movements, cardiac, and muscular) by means of independent component analysis (ICA) led to a significant improvement in detection and localization of frontal and medial temporal sources. We verified our results by using simulations of sources placed in frontal and medial temporal regions with various types of background noise (eye, heart, and visual). We report that the detection and localization accuracy of frontal and medial temporal sources with beamformer techniques is highly dependent on the magnitude and location of background sources and that removing artifacts can substantially improve the beamformer's performance.