Can MEG source modeling resolve laminar brain activity in humans?

Depth inference is a conditional capability, not a default

The question is whether non-invasive magnetoencephalography (MEG), which measures magnetic fields from neuronal currents at millisecond resolution, can tell apart sources sitting at different depths within a cortical column. Previously, human non-invasive work had been limited to coarse deep-versus-superficial distinctions. The paper's answer is that finer depth resolution is achievable, but only inside a narrow operating envelope defined by signal quality, anatomical alignment and forward-model accuracy.

The framework replaces a two-layer deep/superficial model with multiple source surfaces spanning the cortex. In the empirical analyses, the authors generated 9 equidistant intermediate surfaces between the pial boundary and the white matter boundary, giving 11 cortical surfaces in total. These surfaces are explicitly not the six histological laminae; they sample cortical depth the way a laminar electrode with evenly spaced contacts would. Each surface becomes a candidate source layer, and Bayesian model comparison asks which layer best explains the measured field.

That design choice matters for reading every result below. When the paper reports an error in layers or millimetres, it is reporting the distance between the simulated or assumed source depth and the depth of the model with peak evidence. It is not reporting a histological laminin assignment unless the analysis explicitly maps surfaces onto cytoarchitectonic laminae.

Signal-to-noise ratio sets the depth limit

The simulations used synthetic MEG data built from real cortical surfaces and a real sensor layout, with sources placed at 100 random cortical locations. Dipole strength was 10 nAm and the simulated patch had a 5 mm full-width at half-maximum spatial extent, with white noise added at the sensor level to reach target per-trial signal-to-noise ratios (SNR).

Accuracy depended strongly on that SNR. Coarse deep-versus-superficial inference was possible in principle at −50 to −35 dB, while fine-scale laminar inference became possible at −20 dB and above. Reconstruction error fell below chance at −35 dB and improved at higher SNR, and the probability density of the maximum-evidence model showed a diagonal pattern from −35 dB, meaning the peak evidence usually landed on the simulated depth.

The paper also reports a methodological dependency inside this result: at low SNR, inference was biased when using the empirical Bayesian beamformer algorithm rather than the model with a spatial prior used in the main analyses. So the SNR threshold is not a property of MEG alone; it is a property of MEG plus the inversion method. A reader comparing studies should check which estimator produced the reported depth resolution.

Co-registration error above 2 mm erases the advantage

Co-registration is the alignment between the MEG sensor positions and the individual's MRI-defined head and cortical surfaces. The authors simulated misalignment by applying random rigid-body perturbations to the three MRI fiducials, then varied the error from 0 to 5 mm while holding SNR at 0 dB.

At 2 mm or less, the surface model with the highest mean free energy corresponded to the simulated cortical depth, and the peak-evidence probability density was strongly diagonal. At 3 to 5 mm, mean free energy still showed significant diagonality, but the probability density indicated only a gradient between the deepest and most superficial surfaces. Reconstruction error was lowest at 0 mm and rose with greater misalignment, remaining significantly better than chance up to 2 mm but reaching chance level at 3 mm and above. No significant bias appeared at any co-registration level.

The authors note that their perturbation is a controlled measure of co-registration error and may overestimate the effective misalignment relative to more spatially distributed or temporally varying head movement, because realistic head motion and its correction were not explicitly simulated. They therefore describe the reported error levels as conservative upper bounds. Deep sources were less affected by co-registration error than superficial ones, and a coarse deep-versus-superficial distinction may still be possible at some locations with errors up to 5 mm if SNR is sufficient.

Column orientation and lead-field separability decide where it works

A dipole's orientation must be constrained before source reconstruction, and the paper tested seven ways of estimating cortical column orientation, including surface normals from the original high-resolution mesh, normals from a downsampled mesh, cortical patch statistics and link vectors. Link vectors connect corresponding vertices on the pial and white matter surfaces and are equivalent across all intermediate surfaces, which keeps model-based layer comparisons dependent on dipole location rather than on orientation differences between layers.

Getting orientation wrong systematically degraded localization accuracy. The paper states this as a confirmed requirement: accurate laminar inference depends on correctly estimating cortical column orientation, and errors in that estimate reduce accuracy.

Anatomy then modulates how well depths can be separated at all. Regional variation in cortical anatomy influenced reconstruction fidelity, and lead-field separability, the degree to which sources at different depths produce distinguishable sensor patterns, emerged as the strongest predictor of where inference works. Cortical thickness, dipole orientation and scalp-to-cortex distance shaped the spatial distribution of accuracy nonlinearly. When the authors re-analysed simulations using cytoarchitectonic laminae mapped from the BigBrain atlas, thicker laminae showed systematically lower reconstruction error, and middle laminae showed greater variability, consistent with errors that can run in either direction from the middle of the depth axis.

Empirical data recover canonical feedforward and feedback motifs

The framework was then applied to empirical event-related fields from three independent datasets, covering visual and sensorimotor circuits. The test was whether reconstructed laminar patterns matched established physiological expectations, not whether they proved a specific circuit mechanism.

In primary visual cortex during a visuospatial attention task, visually evoked fields showed a feedforward sequence: early dominance of the granular family, consistent with thalamo-cortical input, followed by supragranular dominance at the response peak. The same sequence appeared in both contralateral and ipsilateral hemispheres. In motor circuits, combining two independent datasets, primary motor cortex fields were dominated by infragranular activity beginning before movement onset and recurring around execution, consistent with corticospinal output, followed by delayed supragranular responses. Motor-evoked responses in primary somatosensory cortex showed predominant superficial activity at movement onset, consistent with somatosensory reafference.

The authors interpret the convergence of these patterns across cortical systems, task variants and acquisition sites as evidence that multilayer high-precision MEG can recover established physiological motifs at the level of laminar families under favourable conditions. That is a plausibility claim about the method, supported by agreement with prior physiology, rather than a direct validation against invasive laminar recordings in the same participants.

Dominant laminar contribution, not simultaneous generators

The simulations and the empirical analyses support a narrower claim than full laminar decomposition. The authors state that results are best interpreted as identifying the dominant laminar contribution within a localized cortical region and time window, rather than fully resolving multiple simultaneous generators.

A separate simulation with two equal-strength sources at different depths tested this directly, with one source fixed in the middle layer and the second varied from the most superficial to the deepest surface at −20 dB SNR. The paper's discussion frames the limitation in terms of forward-model accuracy and sensitivity to analysis choices such as patch size, sliding-window duration and preprocessing.

Supplementary robustness simulations reported that the main conclusions were largely preserved under temporally autocorrelated noise, concurrent interfering sources, source-space mismatch and forward-model mismatch, although forward-model mismatch reduced the range of SNRs and co-registration errors over which accurate inference was achieved. The practical reading is that a reported laminar pattern is a statement about which depth dominates a local region in a time window, conditional on the forward model being approximately right.

Millisecond timing is the advantage over laminar fMRI

The paper compares high-precision MEG with ultra-high-field fMRI at 7 to 11.7 Tesla for laminar work. fMRI gives spatially precise laminar profiles but depends on the hemodynamic response, which is subject to vascular drainage effects that can bias profiles toward superficial layers; deconvolution, vascular correction and contrast-based designs improve this but do not fully remove the problem. The paper also cites low temporal resolution for fMRI, around 12 s per whole-brain acquisition at 0.8 mm resolution, which leads many studies to restrict acquisition to part of the brain.

High-precision MEG instead provides millisecond-resolution measurements of neuronal population activity with whole-head coverage and is unaffected by vascular artifacts, which the authors argue makes it suited to rapid laminar dynamics such as beta bursts, induced oscillatory activity and event-related responses. The cost is that MEG needs anatomical imaging to define laminar boundaries, and current analyses rely on a histological atlas based on a single ex vivo brain, the BigBrain atlas, because individualized in vivo laminar boundary estimation across the whole cortex remains a major technical challenge.

The paper also notes that the estimated number of pyramidal cells needed to generate a detectable MEG signal ranges from 10,000 to 50,000, with most of the signal arising from laminae II/III and V, and that neuronal density varies from about 20,000 cells/mm2 in frontal regions to around 45,000 in sensorimotor areas and up to 90,000 in the foveal region of V1. These anatomical constraints sit alongside cortical thickness, gyral folding and column orientation in determining achievable depth precision.

Check SNR, co-registration and orientation before trusting a depth claim

The paper's most decision-relevant output is a threshold structure rather than a single number. Coarse deep-versus-superficial inference can survive SNR as low as −50 to −35 dB, but full laminar inference needs about −20 dB or better. Co-registration error must be about 2 mm or less for accurate laminar reconstruction, with 3 mm and above reaching chance level in the simulations, while a coarse distinction may persist at up to 5 mm in some locations given sufficient SNR. Correct cortical column orientation is a prerequisite, not a refinement, and lead-field separability predicts which cortical locations will work best.

For a reader evaluating a laminar MEG claim, those three variables, plus the inversion method used, are the first things to check, because the paper shows each one can move the result from accurate to chance-level. The empirical feedforward and feedback patterns in visual and sensorimotor cortex are consistent with canonical physiology, but they were obtained under favourable conditions and describe dominant laminar contributions within local regions and time windows. The paper's own forward-looking suggestion is subject-specific reference measurements to quantify achievable SNR and constrain later task analyses, which would turn these group-level feasibility limits into a per-participant reliability check.

Sources used in this answer