Magnetoencephalography (MEG) is a noninvasive method for recording the extremely weak magnetic fields generated by electrical activity in the brain. Arrays of sensitive magnetic detectors measure signals outside the head, providing a millisecond-scale record of neural activity. Mathematical models can then estimate the locations and time courses of the underlying sources. The term magnetic source imaging usually refers more specifically to combining these source estimates with anatomical images.
MEG is important because it measures a direct physical consequence of neural electrical activity rather than the slower changes in blood flow and oxygenation measured by functional magnetic resonance imaging (fMRI). Its principal established clinical application is the evaluation of selected people with epilepsy, particularly when surgery is being considered. It is also used for functional mapping and for research on perception, language, movement, memory, and brain networks.
The method combines unusually demanding magnetic instrumentation with an inherently uncertain reconstruction problem: detecting a brain signal does not uniquely reveal where it originated. Understanding MEG therefore requires attention both to its biological and physical foundations and to the assumptions used in acquisition, analysis, and interpretation.
Historical development
The first widely recognized measurement of the human brain’s magnetic field was reported in 1968 by David Cohen. Working at the University of Illinois, Cohen used a large copper induction coil inside a magnetically shielded room to detect the magnetic counterpart of the alpha rhythm, an approximately 8–13 Hz oscillation prominent during relaxed wakefulness with the eyes closed. The signal was extremely small and required averaging and substantial suppression of environmental interference. This experiment demonstrated that brain activity could be detected magnetically without electrodes contacting the scalp.
A decisive advance followed the introduction of the superconducting quantum interference device (SQUID) into biomagnetic measurement. James Zimmerman and colleagues developed highly sensitive superconducting detectors, and Cohen reported SQUID recordings of brain activity in 1972. These instruments offered much greater sensitivity at low frequencies than ordinary induction coils. Their development linked MEG to advances in low-temperature physics, superconducting electronics, and the measurement of other biological magnetic fields, especially those generated by the heart.
Early MEG systems measured only one or a few positions at a time. Investigators moved the detector across the head to map a response, often repeating the same stimulus many times at each position. This made measurements slow and complicated the study of activity that changed between trials. Multichannel instruments developed during the 1980s, followed by whole-head systems in the 1990s, allowed many regions to be recorded simultaneously. Helmet-shaped sensor arrays made it practical to examine rapidly changing field patterns and to collect clinically useful recordings within a single session.
The integration of MEG with magnetic resonance imaging (MRI) was equally important. MRI supplied individualized anatomical information, allowing estimated magnetic sources to be displayed relative to cortical folds, lesions, and other structures. Improvements in digital computation supported increasingly elaborate head models and source-reconstruction methods. The field consequently moved from identifying simple evoked responses toward reconstructing distributed activity and investigating interactions between brain regions.
From the 2010s onward, optically pumped magnetometers (OPMs) became a major direction of development. These sensors detect magnetic fields through the optical manipulation of atomic spins rather than superconducting circuitry. Their use in wearable arrays, including a prominent wearable MEG demonstration published in 2018, challenged the longstanding requirement that participants remain beneath a fixed cryogenic helmet. This development has not eliminated the need for magnetic shielding or careful modeling, but it has expanded the range of possible sensor geometries and experimental movements.
Neural generators and magnetic-field physics
MEG signals arise mainly from the summed intracellular currents associated with postsynaptic activity in large populations of neurons. Cortical pyramidal neurons are particularly important because their elongated dendrites often have similar orientations within a small patch of cortex. When many such cells are active together, their currents can sum coherently rather than cancel. The resulting magnetic field extends through surrounding tissues and can be detected outside the head. An individual neuron ordinarily produces a signal far below the sensitivity of conventional noninvasive MEG.
The currents are commonly divided into primary currents, representing the neural generators, and secondary or volume currents flowing through conductive tissue. Both contribute to the electromagnetic field, although their relative effects depend on anatomy and measurement geometry. The cerebral cortex is folded, so neighboring populations can point in substantially different directions. Activity in opposing banks of a sulcus may partially cancel even when both populations are strongly active. Consequently, measured signal strength is not a simple measure of how many neurons are active.
At frequencies ordinarily studied with MEG, the electromagnetic problem is treated using a quasi-static approximation. Propagation delays across the head are negligible for the relevant biological timescales. A small active cortical region is often approximated as a current dipole, characterized by a location, orientation, and moment. In a homogeneous, unbounded conductor, its field can be expressed as
where is the current-dipole moment and its position. Realistic head calculations must also account for conductivity boundaries and volume currents; this expression alone is not a complete head model.
Brain magnetic signals commonly fall in the range of tens to hundreds of femtotesla, although amplitudes vary substantially with the source and recording conditions. A femtotesla is tesla. Earth’s magnetic field is of the order of tens of microtesla, hundreds of millions of times larger than many MEG signals. This disparity explains why MEG requires extraordinary control of background fields and interference.
MEG’s sensitivity also depends strongly on source orientation and depth. In an ideal spherical conductor, an exactly radial current dipole produces no external magnetic field; tangential sources are detectable. Real head geometry makes this distinction less absolute, but it remains practically important. Superficial cortical sources are generally easier to detect than equally strong deep sources. Unlike scalp electrical potentials, magnetic fields are less distorted by differences in skull conductivity, but MEG is not independent of anatomy, conductivity, or source cancellation.
Sensors, shielding, and instrument design
Most established whole-head MEG systems use SQUID sensors cooled to approximately 4 kelvin with liquid helium or related cryogenic arrangements. Magnetic flux collected by superconducting pickup coils is coupled to the SQUID, whose response is maintained within a useful operating range by feedback electronics. The sensors sit inside a thermally insulated vessel called a Dewar. Its inner surface follows the approximate shape of the head, but insulation and the vacuum space keep the sensing coils some distance from the scalp.
Sensor arrays use different measurement geometries. A magnetometer measures a component of the magnetic field, whereas a gradiometer measures how a field component changes across a spatial separation. Distant interference often varies relatively little across a short baseline, so differencing can reduce its contribution while retaining more of the rapidly varying field from nearby brain sources. Axial and planar gradiometers have different spatial response patterns. One widely used whole-head configuration has 306 measurement channels at 102 positions, combining a magnetometer with two planar gradiometers at each position; other systems use different channel counts and arrangements.
Recordings usually take place in a magnetically shielded room. High-permeability metal layers divert low-frequency magnetic flux, while conductive layers help attenuate time-varying interference through induced currents. Shielding does not make the environment perfectly field-free. Power-line fields, elevators, vehicles, nearby machinery, and magnetic contamination can remain relevant. Reference sensors, active field compensation, and computational separation of internal and external field components provide additional suppression.
OPMs use a different physical principle. Light prepares and probes the spin state of an atomic vapor, often an alkali metal such as rubidium. Changes in spin behavior reveal the local magnetic field. The vapor cell is heated, but insulation allows a packaged sensor to operate close to the scalp. Unlike a SQUID array, an OPM array does not require a shared liquid-helium vessel, permitting sensors to be mounted in individualized or adjustable headgear. Reduced source-to-sensor distance can increase measured signal strength, particularly when compared with a poorly fitting fixed helmet.
These advantages come with engineering constraints. Many highly sensitive OPMs operate within a narrow range around a very low background field, so residual fields must be carefully nulled. Movement through remaining field gradients can create artifacts much larger than brain signals. Sensor cross-talk, calibration, thermal management, and precise knowledge of sensor positions also require control. Measuring multiple field components can improve the information available, but comparisons between instruments must consider more than channel count: noise level, placement, bandwidth, field stability, and calibration all influence performance.
Recording procedures and signal preparation
A MEG examination begins with a protocol matched to the scientific or clinical question. Recordings may capture spontaneous activity at rest, responses to sensory stimulation, performance of a language or movement task, or epileptiform discharges. MEG itself neither delivers ionizing radiation nor applies the strong magnetic fields used in MRI. Nevertheless, nearby metal and magnetized objects can interfere with recording, and the full examination environment—including associated MRI or sedation, if used—has its own safety procedures.
The position of the head relative to the sensor array must be established. In conventional systems, small coils attached to the head generate identifiable signals that permit localization within the helmet. A digitizer records the positions of these coils, anatomical landmarks, and often many points on the scalp. These measurements connect the MEG coordinate system to the person’s structural MRI. Continuous or repeated head-position measurements can document movement during acquisition. In wearable systems, the geometry between sensors and head may be relatively stable, but its accuracy still has to be measured.
MEG is often acquired together with electroencephalography (EEG), as the two modalities have complementary sensitivities. Additional recordings may monitor eye movements, cardiac activity, and muscle activity. Eye blinks, jaw tension, swallowing, heartbeat, and small movements of magnetic material can all generate artifacts. Dental hardware, implanted devices, or residual magnetization may impair particular recordings without necessarily making every channel unusable. The importance of an artifact depends on its strength, location, frequency content, and relationship to the signal being studied.
Sampling rates are commonly around 1,000 samples per second or higher, although the appropriate rate depends on the desired bandwidth. Stimulus timing must be measured accurately: a delay between a computer command and the actual onset of a sound or image can shift the apparent latency of a neural response. Experimental recordings therefore often include physical measurements of stimulus delivery. Quiet intervals recorded without a participant can help characterize instrument and environmental noise.
Preprocessing may include identifying malfunctioning channels, excluding contaminated intervals, filtering, correcting head movement, and separating interference from brain signals. Independent component analysis can help identify components associated with heartbeat or blinking, while signal-space methods exploit spatial patterns or physical field properties. Each operation can also remove or distort neural activity if its assumptions are inappropriate. Filtering can alter waveform shape, and aggressive artifact rejection can selectively discard meaningful behavior.
For event-related studies, repeated trials are often averaged to emphasize activity consistently timed to a stimulus. Ideally, uncorrelated noise decreases approximately with the square root of the number of trials. Activity whose timing or phase varies across trials may disappear from the average while remaining detectable in single-trial power estimates. Resting-state analysis uses different strategies, usually emphasizing oscillatory spectra and relationships between regions rather than a response locked to an external event.
Source reconstruction and statistical inference
MEG sensors measure magnetic fields, not anatomical source locations. The inverse problem is to infer the neural currents that could have generated those fields. It is nonunique: different current distributions can produce indistinguishable measurements. A source image therefore depends on a combination of observations, a physical model, anatomical information, and explicit or implicit assumptions. A visually precise map is not evidence that those assumptions have yielded an equally precise biological localization.
The corresponding forward problem predicts sensor signals from a specified source. Structural MRI can define the scalp, skull, brain surface, and possible source positions. Depending on the application, the conductor may be represented by a sphere, a set of boundary surfaces, or a numerical volume model. The relationship is often written
where contains the measured signals, represents candidate source currents, is the lead-field matrix, and represents noise and model discrepancy. Errors in sensor positions, MRI alignment, or head geometry propagate into the inferred sources.
Equivalent-current-dipole fitting assumes that one or a few compact generators explain a selected field pattern. The algorithm adjusts their locations, orientations, and strengths to match the observations. This approach is widely used for well-defined epileptiform spikes and certain evoked responses. It is most interpretable when the assumed small number of generators is plausible. A high goodness of fit does not establish that a fitted dipole is the unique source, and a dipole can represent the net effect of activity distributed over a cortical patch.
Distributed methods instead place many candidate sources throughout a cortical surface or brain volume. Minimum-norm estimation favors a solution with relatively small overall source amplitude, while depth weighting and noise normalization modify characteristic biases. Regularization stabilizes the solution by imposing constraints or penalties. Bayesian approaches express assumptions through prior distributions and estimate source properties conditionally on those assumptions. No single choice is optimal for every combination of source extent, signal strength, and noise.
Beamforming constructs spatial filters designed to estimate activity from a target location while suppressing contributions from elsewhere. It is useful for mapping oscillatory power and some event-related responses, but highly correlated sources can be difficult to separate. Source reconstructions are also affected by leakage: activity from one generator can spread into estimates assigned to neighboring or distant locations.
Statistical inference adds another layer. Testing many sensors, time points, frequencies, or source locations creates a multiple-comparison problem. Permutation tests and cluster-based procedures can address particular forms of this problem, but a significant cluster does not automatically pinpoint the exact onset or boundary of an effect. Localization accuracy is therefore conditional rather than fixed; it varies with signal quality, source configuration, model accuracy, and the question being tested.
Clinical applications in epilepsy and functional mapping
MEG’s principal established clinical role is as an additional source of information in the presurgical evaluation of selected patients with drug-resistant focal epilepsy. A central question is where clinically important epileptic activity originates and how that region relates to tissue essential for movement, sensation, or language. MEG does not answer this question alone. Its findings are interpreted alongside seizure history, scalp video-EEG, structural imaging, and, where relevant, other functional imaging and intracranial recordings.
Most clinical epilepsy MEG examinations analyze interictal epileptiform discharges: brief abnormal events occurring between seizures. The magnetic field around a spike can be modeled using one or more equivalent current dipoles or other reconstruction methods. Repeated discharges may produce a cluster of estimated sources, helping define a candidate region for further investigation. Magnetic source imaging is particularly useful when it supplies localization information that is absent, ambiguous, or discordant in other tests, including some cases without an obvious structural lesion on MRI.
An essential distinction separates the irritative zone, which generates interictal discharges, from the seizure-onset zone and the broader epileptogenic zone. These regions may overlap but are not identical. A cluster of spike sources therefore does not by itself define tissue whose removal would stop seizures. Activity can propagate rapidly, and the most prominent part of a recorded spike may reflect a propagated response rather than its earliest generator. Simultaneous EEG can help characterize events with different electrical and magnetic visibility.
MEG findings can contribute to planning where intracranial electrodes are placed and to evaluating whether proposed anatomical hypotheses are consistent with the observed physiology. They may also influence the interpretation of a subtle lesion or the boundaries of an area requiring further investigation. Clinical outcome studies require careful interpretation because patients undergoing MEG are often selected for difficult localization problems, and results may alter subsequent testing and treatment. Surgical outcome cannot be attributed to a single presurgical test without accounting for that broader pathway.
Functional mapping provides another clinical use. Somatosensory responses to peripheral nerve stimulation can help localize sensory cortex, and movement-related activity can help identify motor regions. Auditory and visual responses provide information about corresponding sensory systems. Language tasks can estimate the distribution and hemispheric predominance of language-related activity, although results depend on task design, analysis method, and the person’s performance. Such maps can inform neurosurgical planning but are not interchangeable with every form of invasive functional mapping.
Clinical reports distinguish recorded observations from their interpretation. Relevant features include discharge timing and morphology, the number and consistency of localized events, source distribution, fit quality, and concordance with other findings. An examination in which no epileptiform discharges are captured does not exclude epilepsy, and an interpretable magnetic abnormality does not independently establish its clinical significance.
Cognitive neuroscience, oscillations, and brain networks
MEG is especially valuable for studying the timing of neural processing. Auditory responses include a prominent cortical component near 100 milliseconds after sound onset in many adult experimental conditions. Visual, somatosensory, and language-related tasks produce their own sequences of responses. By comparing their timing and source estimates, researchers investigate how information moves through sensory and association systems. Latencies and amplitudes depend on factors such as stimulus properties, attention, age, and recording conditions rather than being fixed signatures of a single mental operation.
Analysis of neural oscillations extends beyond averaged evoked responses. Commonly studied bands include delta, theta, alpha, beta, and gamma, although their boundaries vary across studies. Alpha activity often changes with visual input and attention; sensorimotor beta activity typically decreases during movement and may increase afterward. Gamma-band activity can accompany sensory and cognitive processing, but high-frequency recordings are particularly vulnerable to muscle contamination. A frequency label identifies a range of activity, not a unique biological function.
Time–frequency methods examine how oscillatory amplitude or phase changes over the course of a task. They distinguish responses that are consistently phase-locked to an event from changes in power that occur with variable phase. This distinction matters because a clear increase in single-trial oscillatory power can coexist with little change in the trial-averaged waveform. Conversely, a transient evoked response can generate broad spectral changes without representing a sustained oscillation. Interpretation requires considering both the waveform and its spectral representation.
MEG also supports the study of functional connectivity, meaning statistical relationships between activities in different regions. Measures include correlations between amplitude envelopes, phase relationships, coherence, and model-based estimates of directed interactions. However, common input and source leakage can create apparent coupling without direct communication between the reconstructed regions. Some methods reduce sensitivity to instantaneous shared components, but this can also discard genuine interactions. A statistical direction of influence is not, by itself, proof of anatomical transmission or biological causation.
Research combines MEG with structural MRI, fMRI, and diffusion MRI to relate rapid activity to anatomy, slower hemodynamic responses, and white-matter pathways. Comparisons with invasive recordings help evaluate what noninvasive signals represent. Machine-learning methods can decode stimulus categories or behavioral variables from spatial and temporal patterns, but successful prediction does not uniquely identify the neural mechanism producing the information.
MEG has also been investigated in conditions including schizophrenia, autism spectrum disorder, and Alzheimer’s disease. Studies examine features such as sensory responses, oscillatory changes, and network organization. Differences between research groups do not necessarily provide an accurate individual-level biomarker. Clinical translation requires reproducibility, appropriate comparison populations, and evidence that a measurement contributes useful information beyond existing assessments.
Limitations, reproducibility, and emerging directions
MEG’s practical limitations begin with the small size of the signal. Environmental interference, physiological artifacts, and head movement can obscure neural activity even when the sensors themselves are highly sensitive. Conventional cryogenic systems also impose geometrical constraints: a fixed helmet fits some heads better than others, and greater sensor-to-scalp distance weakens the signal. This is particularly relevant for infants and young children. Shielded rooms, specialized instruments, maintenance, and expert analysis contribute to costs and restrict availability.
Biological visibility is uneven. Favorable superficial sources can be measured clearly, while deep, weak, spatially extended, or mutually cancelling activity may be difficult to recover. Studies have reported MEG sensitivity to structures such as the hippocampus, but deep-source claims require particular care because stronger cortical sources and reconstruction assumptions can produce misleading localizations. Failure to detect a response does not establish the absence of neural activity, and differences in measured amplitude can reflect anatomy or geometry as well as physiology.
Comparisons with EEG and fMRI illustrate why no modality is a universal substitute for the others. EEG is generally less expensive and easier to deploy, while having different sensitivities to source orientation and head conductivity. fMRI offers extensive spatial coverage through an indirect hemodynamic signal but does not ordinarily resolve neural events on MEG’s millisecond timescale. Combining modalities can constrain interpretations, yet their signals are not interchangeable measurements of a single quantity. Discordant findings can arise from genuinely different biological sensitivities.
Reproducibility depends on details that extend well beyond the reconstruction algorithm. Sensor calibration, coordinate conventions, head position, trial exclusion, filtering, noise estimation, and statistical choices all affect results. The Brain Imaging Data Structure has been extended to MEG to standardize the organization of recordings and metadata. Widely used open-source analysis packages include MNE-Python, FieldTrip, and Brainstorm. Such tools facilitate scrutiny and reuse, but reproducible software execution does not guarantee that an experimental inference is valid.
Wearable OPM arrays are a major emerging direction because they can accommodate different head sizes and permit more natural movement. Potential applications include developmental studies and tasks involving interaction or movement that are difficult under a fixed helmet. Larger arrays, improved magnetic compensation, reliable sensor localization, and stable calibration are continuing engineering priorities. Movement tolerance remains conditional on the field environment rather than an intrinsic guarantee of the sensor technology.
Another priority is stronger validation of clinical and computational claims. Repeated measurements, physical test sources, realistic simulations, comparisons with intracranial recordings, and independent clinical cohorts address different aspects of validity. A method that accurately reconstructs a simulated isolated source may perform differently with simultaneous biological generators. Likewise, a classifier that separates selected research groups may fail in a clinically representative population. Establishing utility requires evaluation under the conditions in which the measurement is intended to be interpreted.