Magnetoencephalography Without a Shielded Room

Note up front: only the abstract of this paper is publicly available to me, so this explainer builds the necessary background and physics from general knowledge, is explicit about which claims come from the paper, and flags exactly what to verify once the full text is in hand.

TL;DR: A UCL/CEA team reports a multichannel MEG system that works with no magnetically shielded room at all — the entire apparatus weighs under 75 kg, versus the multi-tonne, million-dollar shielded rooms that every MEG system on Earth currently requires. They claim in-vivo signal-to-noise comparable to a conventional cryogenic MEG inside a shielded room, achieved by combining low-cost active shielding (coils that null the ambient field) with software spatial filtering (exploiting the fact that brain signals and environmental interference have different spatial fingerprints across a sensor array). If the quantitative results hold up, this removes the single largest infrastructure barrier to millisecond-resolution, non-invasive human neuroimaging — with obvious consequences for clinical neurology, BCI research, and anyone who has ever wanted MEG outside a purpose-built basement.

Why the shielded room is the whole problem

MEG measures the magnetic fields produced by neuronal currents — mostly synchronized postsynaptic currents in cortical pyramidal cells. Unlike EEG, magnetic fields pass through skull and scalp essentially undistorted, so MEG gives you EEG's millisecond temporal resolution with meaningfully better spatial localization. It is, in principle, one of the best non-invasive windows we have onto fast neural computation.

The catch is dynamic range. Evoked brain fields at the scalp are on the order of tens to hundreds of femtotesla (10^{-13} T). The Earth's static field is about 5 \times 10^{-5} T. Urban magnetic noise — elevators, trams, power lines, passing cars — fluctuates at nanotesla-to-microtesla scales. The paper's headline framing is that they localize brain signals over 300 million times smaller than environmental interference. Sanity check: 3 \times 10^8 \times 100\,\text{fT} = 30\,\mu\text{T} — that's the ambient field scale, so the claim is that they're operating fully exposed to Earth-scale magnetic environments, not just a noisy lab.

The traditional solution is brute force: a walk-in room built from multiple nested layers of mu-metal and aluminum, weighing several tonnes and typically costing on the order of a million dollars installed, often with structural building modifications. This — not the sensors — is why there are only a few hundred MEG systems worldwide, concentrated in wealthy research hospitals. The paper's comparison: their whole system is under 75 kg, "more than 100 times lighter than a typical shielded room."

There's a second, subtler reason shielding has been hard to drop. The wearable-MEG revolution of the last decade was built on optically pumped magnetometers (OPMs) — quantum sensors using alkali vapor cells — but the standard high-sensitivity variety operates in the SERF regime, which only works near zero absolute field. So the shielded room wasn't just noise suppression; it was a precondition for the sensors functioning at all. Any credible "no shielded room" system has to solve both problems: keep the sensors in their operating regime, and dig a femtotesla signal out of microtesla-scale, time-varying interference. (The abstract doesn't specify the sensor technology. Informed speculation, clearly flagged as such: the author list includes CEA researchers known for helium-4 OPMs, which have far wider dynamic range than SERF sensors and have previously been demonstrated in weakly shielded settings — a natural fit for this design. Verify against the methods.)

The key idea: interference and brain signals live in different spatial subspaces

The system, per the abstract, rests on two components: low-cost active shielding and software-based spatial filtering. Neither is new in isolation. The claim is that together, with a multichannel array, they're sufficient to replace the room entirely.

Active shielding is the noise-cancelling-headphones move: reference measurements of the ambient field drive current through coils arranged around the sensor array, generating an equal-and-opposite field. This does two jobs at once — it nulls the large quasi-static field (keeping sensors in range and within their operating regime) and it attenuates low-frequency environmental fluctuations. Coils and current drivers are cheap and light; this is where most of the 100× weight saving comes from. But feedback nulling is imperfect: it's bandwidth-limited, and a small number of coils can't cancel a spatially complex interference field at every sensor simultaneously. Residual interference after active shielding will still dwarf brain signals.

Spatial filtering is where the residual gets killed, and it's the conceptually beautiful part. The physics: a magnetic source at distance d produces a field whose variation across a sensor array of size L scales roughly like L/d. A tram 100 m away produces a field that is, across a 20 cm head-mounted array, uniform to a fraction of a percent — every sensor sees essentially the same vector, wobbling in time. A cortical current source 3 cm beneath the sensors produces a sharply structured dipolar pattern that changes sign across the array. Interference is enormous but spatially boring; brain signals are tiny but spatially rich.

Concretely, one models the field across the array with a spatial basis — in the simplest case (homogeneous field correction, a method this group pioneered) a uniform field plus gradients; more generally a low-order harmonic expansion, as in the signal-space separation methods used in commercial MEG. External sources project almost entirely onto the low-order terms; brain sources have substantial energy in higher-order, near-field components. Regress out the low-order subspace at every timepoint and you get orders of magnitude of interference rejection in software — if your sensor array is dense and well-calibrated enough to estimate the smooth component accurately. This is why the abstract stresses "multichannel": one magnetometer in an unshielded room measures the tram. Dozens of magnetometers measure the tram and its spatial signature, which is exactly what lets you delete it. The multichannel array is simultaneously the interference canceller and the source-localization instrument.

So the architecture, as best one can reconstruct from the abstract: coils null the bulk field so the sensors can operate at all and take out the largest fluctuations in analog; the multichannel array plus spatial filtering removes what's left in software; and the surviving high-spatial-order data feeds standard MEG source localization. Three stages, each cheap, each attacking a different part of the 8–9 orders of magnitude.

What the paper claims to show — and what to check

From the abstract, the evidence is: (1) in-vivo recordings with SNR "comparable to what can be obtained from a conventional cryogenically-cooled MEG system sited within a shielded room"; (2) not just detection but localization of brain activity without passive shielding; (3) the full system at <75 kg.

That's a strong headline and a thin evidentiary summary, so here is precisely what to scrutinize in the full text:

The SNR comparison. "Comparable SNR" needs numbers: which paradigm (evoked responses like auditory M100? alpha rhythm? somatosensory?), how many trials averaged, what frequency band. Unshielded operation is hardest at low frequencies (< a few Hz), where environmental drift is worst and where a lot of clinically interesting activity lives (slow-wave sleep, epileptiform slowing, movement-related fields). An SNR match on a 10–40 Hz evoked response would be impressive but would not establish that the low-frequency band is usable. Check the noise spectra.

The environment. "Without a shielded room" spans a huge range — a magnetically quiet rural lab is a very different beast from a hospital floor with elevators and an MRI down the hall. The whole promise here is deployment in realistic environments; the demonstration environment is the crux.

Localization accuracy. The abstract claims localization but quotes no error. For this to substitute for conventional MEG (e.g., presurgical epilepsy mapping), source localization needs to land within roughly a centimeter of ground truth (phantom tests or comparison against the same subject in a cryogenic system). Look for a phantom experiment.

Movement and robustness. A signature advantage of wearable OPM-MEG is tolerating head movement. But movement in a large background field converts static field structure into sensor-frame signal — the harder the environment, the harder motion is. Was the subject rigidly still? Spatial filtering also removes some brain signal (the projection isn't perfectly orthogonal to neural sources); the signal-loss cost should be quantified.

Channel count and calibration. Harmonic spatial filtering degrades sharply with sensor gain/orientation miscalibration — at these interference levels, a 0.1% gain error on a microtesla-scale residual is a nanotesla artifact, still ~10⁴× the brain signal. How they achieve and maintain calibration outside a shielded environment may be the most technically interesting part of the methods.

None of these caveats is a reason to dismiss the result; this group (Tierney, Barnes, Mellor at UCL) has an extensive, credible track record in OPM-MEG interference rejection, and the trajectory of their prior work — from shielded rooms, to lightly shielded operation, to this — makes the claim plausible rather than shocking. But "comparable SNR" is doing a lot of work in that abstract, and the difference between "works in our courtyard" and "works in a clinic" is the entire value proposition.

Why this matters if it holds

The comparison to keep in mind is what happened when MRI didn't need superconducting infrastructure: ultra-low-field portable MRI (e.g., Hyperfine) traded image quality for the ability to exist at a bedside in a rural hospital, and found genuine clinical niches. MEG's situation is more extreme — the room is a larger fraction of the cost and siting burden than the magnet is for MRI — and unlike low-field MRI, the claim here is no quality sacrifice: SNR parity with cryogenic systems.

A 75 kg, room-free MEG changes who can use the modality. Epilepsy presurgical mapping — MEG's main clinical use — could move from a handful of tertiary centers to ordinary hospitals. Pediatric and developmental imaging becomes far more practical (wearable OPM arrays already tolerate movement better than fixed cryogenic helmets; removing the room removes the remaining constraint). For the BCI-minded reader: this puts a non-invasive modality with genuine millisecond resolution and cortical-source localization into the "could be deployed in a normal lab, or eventually an office" category — something EEG offers only with much blurrier spatial information. And for basic neuroscience, MEG during naturalistic behavior stops being gated by the geometry of a metal box.

There's also a quieter methodological point worth appreciating: this is a case study in replacing passive physical infrastructure with sensing plus computation. Eight-plus orders of magnitude of interference rejection, previously bought with tonnes of mu-metal, bought instead with coils, a dense sensor array, and a projection onto the right subspace. That pattern — measure the nuisance well enough to model it, then subtract it in software — is the same one that powers adaptive optics and multi-microphone speech enhancement, and it tends to win over time because sensors and compute get cheaper while metal doesn't.

Where to spend your time in the full paper: the methods sections on the active shielding design and the spatial filtering model (basis choice, calibration procedure), and whichever figure shows the noise spectrum and SNR comparison against the cryogenic system — that single figure is where the headline claim lives or dies. Everything above the methods is, given the abstract's framing, likely to be vision; the vision is justified only if that figure is honest about the low-frequency band and the recording environment.