Millisecond-resolution magnetic fields from cortical currents — exquisite timing, and the magnetic skull is effectively transparent.
MEG sensors detect the magnetic field generated by the same intracellular postsynaptic currents that produce scalp EEG, but from a different component of the current geometry. Cerebral blood flow, skull movement, and other slow sources contribute negligibly, so the signal is essentially pure neurophysiology rather than a mixture requiring as much physiological noise modelling.
The amplitude is the whole problem. Typical brain fields are 10–500 femtotesla — roughly a million times smaller than the Earth's magnetic field and near the noise floor of the room itself. A shielded room, a degaussing system, and cryogenic hardware are not optional extras; they are what makes the measurement possible at all.
This is MEG's central advantage. Electrical potentials are volume-conducted: they smear as they pass through low-conductivity skull and high-conductivity CSF, which is why scalp EEG has no principled inverse solution. Magnetic fields are affected by magnetic permeability, not conductivity, and bone, CSF, and brain are all nearly identical to air in that respect.
The consequence: the field measured outside the head is a faithful sum of the sources inside, with far less spatial distortion. Combined with the fact that a dipole field decays as the inverse cube of distance, MEG is markedly more sensitive to deep sources than EEG — thalamic and hippocampal generators are detectable in MEG but heavily attenuated at the scalp.
Conventional systems use SQUID (superconducting quantum interference device) flux transformers, cooled to around 4 K in a liquid-helium cryostat. Their weakness is geometry: a helmet-sized dewar of liquid helium must be refilled regularly, and the subject must keep still, because the sensors cannot follow the head.
The reason SQUID works so well is gradiometry. Each sensor measures the difference in field between two points ~4 cm apart, so uniform ambient field — the Earth, the building, the building's elevators — cancels exactly. One reference channel is also measured inside the shield to model and subtract whatever common-mode field remains.
Modern systems run 300+ channels, sample above 1 kHz, and restrict analysis to a usable band of roughly 0.01–200 Hz. That band is not a hardware limit but a noise limit: ambient field and thermal noise rise sharply at low frequency.
Because whole-head coverage makes an intact-helmet assumption possible, MEG can be reconstructed without choosing a reference at all. This removes the arbitrary montage decision that complicates scalp EEG, though the cost is a distributed, spatially smeared point-spread function whose extent grows with depth — a trade, not a free win.
The practical families: beamforming (LCMV, DICS) works well for well-separated sources but is biased when sources overlap; beamformers constrained by MRI (sLORETA, eLORETA, MNE) improve spatial precision; and beamforming with parametric empirical Bayes is the current accuracy leader for evoked responses. Combining MEG with scalp EEG, which has complementary sensitivity to different source orientations, is standard in modern pipelines.