MEG Adaptive Noise Suppression using Fast LMS

Magnetoencephalography (MEG) measures magnetic fields generated by electric currents in the brain, non-invasively and with millisecond temporal resolution. Typical signals are 10^-13 T, so noise contamination due to external magnetic fields is a serious concern. Digital signal processing is typically required in addition to magnetic shielding. Using three reference channels, displaced from the head, to measure the noise, we apply adaptive filtering to subtract out estimates of the noise, via the block least-mean-square ("fast LMS") method. The algorithm is tested by its effects on the number and distribution of channels which have statistically significant signals (distinguishable from background noise at a specified false-positive rate). We show that fast LMS both increases the number significant channels and reduces the variance of false positives