Validation of Finite Element Based Forward Approaches Using Anisotropic Volume

The excellent temporal resolution of magnetoencephalography (MEG) is universally acknowledged, but the accuracy of its source estimation is often questioned on theoretical and practical grounds. Here, we evaluate the localization accuracy of MEG for four widely used source modeling techniques (Equivalent Current Dipole (ECD), Synthetic Aperture Magnetometry (SAM), MUltiple SIgnal Classification (MUSIC) and Magnetic Field Tomography (MFT)). A realistically shaped phantom was used, with multiple dipolar sources implanted at superficial as well as deep locations which were driven either separately or simultaneously by weak, transient currents of varied strengths. We also analyze human MEG data elicited by median nerve stimulation experiment. Each technique was tested using single and multiple spheres. High localization accuracy (2-3 mm) was obtained for superficial sources, with all modeling techniques, even when a small portion of trials was used. For the other cases only some of the methods produced acceptable results with only MFT providing consistently localization accuracy of a few millimeters, except for the most weak and deep sources. ECD and MUSIC worked better for sources away from places with rapid shape distortion of the high electrical resistance boundary. Use of multiple spheres produced better results, but the improvement was substantial only for SAM and only for deep sources for ECD. When multiple sources were simultaneously activated, the localization accuracy of SAM and MUSIC dropped sharply, whereas MFT and ECD maintained acceptable (less than 7 mm) accuracy levels. For the human data all methods produced very similar results for the early generators in the primary somatosensory cortex but different results for the stronger components at late latencies.