Title: Blind Separation and Fast Localization for MEG

Barak Pearlmutter

Brain and Computation Lab, University of New Mexico

Abstract:

The strength of magnetoencephalography (MEG) is that magnetic fields are not smeared by the skull, leading to greater potential spatial sensitivity. Unfortunately, this penetrative ability of magnetic fields also causes each sensor to receive signals from a large number of active sources and makes it difficult to shield out noise, both external (power grid) and internal (heartbeat, eyeblinks). We have made a systematic effort to improve MEG’s performance through new signal processing for separation and dipole localization, both raising the effective performance of current MEG systems and relaxing constraints that have constrained MEG hardware design.

For separation, we applied SOBI to complex MEG data. This segregated non-neuronal sources from neuronal ones, and neuronal ones from each other. The separated neuronal sources seem focal, and have temporal responses consistent with their estimated locations. We can routinely isolate sources within modalities, across modalities, across tasks, and across subjects. The SNR is high enough to allow estimation of response onset times for single trials, which can be measured in visual, auditory, and somatosensory modalities with detection rates over 95%. Combined with an improved ability to localize the underlying neuronal sources, this makes possible the non-invasive study of a range of perceptual and memory functions that depend upon the timing of neuronal events occurring in specific brain regions.

With many recovered sources, dipole localization becomes a bottleneck. For both a 4D Neuroimaging Neuromag-122 and the experimental LANL SiS Mark I with superconducting magnetic mirrors, we addressed the single dipole localization problem at low SNR, achieving a reduction in localization time from 449ms to 0.5ms at an accuracy of 12mm, and to 35ms at an accuracy of 2.8mm. Our fast fully automatic noise-robust localizer is suitable for real-time applications, such as closed-loop experimental protocols and brain-computer interfaces.