Advertisement

MEG as an Enabling Tool in Neuroscience: Transcending Boundaries with New Analysis Methods and Devices

  • M. S. HämäläinenEmail author
  • D. Lundqvist
Living reference work entry

Abstract

Neuroscience studies have provided highly detailed information about the anatomical and structural composition and organization of the brain, but insights into its functional principles are still lacking. To advance this understanding, scientists need to match their new research questions with instruments that provide information about the brain’s functionality at a sufficient level of detail with a potential to resolve the very rapidly evolving patterns of brain activity while also providing the necessary spatial detail and accuracy. Until 50 years ago, electroencephalography (EEG) was the only noninvasive technique capable of directly measuring neuronal activity with a millisecond time resolution. However, with the birth of magnetoencephalography (MEG), functional brain activity can now be resolved with this time resolution at a new level of spatial detail.

The use of MEG in practical studies began with the first real-time measurements in the beginning of the 1970s. During the following decade, multichannel MEG systems were developed in parallel with both investigations of normal brain activity and clinical studies, especially in epileptic patients. The first whole-head MEG system with more than 100 channels was introduced in 1992. By the end of the century, hundreds of such instruments had been delivered to researchers and clinicians worldwide. With vibrant interaction between neuroscientists, clinicians, physicists, mathematicians, and engineers, the experimental approaches and analysis methods were developed to establish MEG as an important method to study healthy and diseased brains. With the advent of low-noise room-temperature magnetic field sensors and novel analysis approaches, we are now at the verge of a revolution that will critically improve both the sensitivity and the spatial resolution of MEG and will especially advance its use in studies of early brain development and neurodegenerative disorders, as well as investigations of brain function in naturalistic situations and during interpersonal interactions. This chapter focuses on instrumentation and analysis tool developments, which have enabled and continue to enable MEG to flourish as a noninvasive tool to study brain function. The final section of this chapter offers lessons learned from seasoned investigators on conducting successful MEG studies, necessarily emphasizing additional issues such as the formulation of the research question and creation of experimental protocols.

Keywords

MEG EEG Source estimation MEG instruments Experimental design 

1 Introduction

The structure of the brain can be studied in detail by multiple methods from autopsy specimens or in vivo with magnetic resonance imaging (MRI) and computerized tomography (CT). However, to advance our understanding of functional organization of the human brain and its cognitive capabilities, we need to be able to chart its fine-grained functional and structural organization resolved across space, time, and neuronal oscillation frequency. To advance the understanding of brain function, we thus need to push the technical and methodological boundaries forward, hand in hand with evolution of theoretical models about the brain.

In the study of brain function, the methods of magnetoencephalography (MEG) and electroencephalography (EEG) hold the unique position of being the only noninvasive techniques capable of directly measuring neuronal activity with a millisecond time resolution. These two methods are thus well suited for elucidating the spatiotemporal sequences of brain activity that compose different brain functions while also capturing fundamental brain rhythms and frequency-dependent interactions between different regions.

Whereas the first EEG studies were conducted in the 1920s and 1930s, the first MEG studies began only about 40 years later, using signal averaging with an EEG reference (Cohen 1968). Real-time MEG measurements were enabled in the beginning of the 1970s with the invention and development of the ultrasensitive SQUID sensors. The first recordings made by David Cohen at MIT showed that high-quality recording of the magnetic counterpart of the visual alpha rhythm was possible (Cohen 1972). During the first decade of MEG, pioneering groups recorded many sensory signals and tediously mapped their spatial distributions with single-site devices (Williamson and Kaufman 1981). At the same time, the theoretical foundations of MEG were established, and at the end of the decade, the first equivalent current dipole (ECD) fitting software was created and marked the start of inverse modeling of MEG signals (Tuomisto et al. 1983). By the beginning of the 1980s, it had become clear that for practical neuroscience and clinical studies, multichannel MEG systems ultimately covering the entire scalp would be a necessity. The first whole-head MEG system with more than 100 channels was introduced in Knuutila et al. (1991). By the end of the century, over 200 such instruments had been delivered to researchers and clinicians worldwide. With neuroscientists, clinicians, physicists, mathematicians, and engineers working together in multidisciplinary teams, the neuroscience, clinical applications and analysis methods were developed to establish MEG as an important method to study both healthy and diseased brains.

In this chapter we will review the current state of MEG as a neuroscience research tool and discuss recent developments in instrumentation, analytical techniques, and experimental paradigms. We will also at many points consider the relationship of MEG and EEG, which has over decades intrigued both developers of methods and experimentalists. We will consider the process of seeking answers to neuroscience questions with MEG as a whole. We emphasize the best choices for experimental settings, source estimation methods, and statistical analysis tools, all of which should not be conceived of in isolation but rather as subserving the collective goal of understanding brain function.

2 Information Conveyed by MEG and Its Virtues

The field of neuroscience is experiencing an enormous growth in the ability to record from, and manipulate, brain circuits in humans and in animal models. The different methods are often classified by their capability to discern the temporal details of activity and their ability to precisely and correctly locate the sites of activity and to relate them to the underlying neuroanatomy (see Fig. 1). The multitude of studies employing hemodynamic techniques since the advent of fMRI 25 years ago has perhaps led to an overemphasis on mapping brain function, paying relatively little attention to resolving the underlying spatiotemporal characteristics of the physiological or physical quantity mapped.
Fig. 1

Spatial and temporal resolutions of neuroscience tools: optical imaging. EEG electroencephalography, fMRI functional MRI, MEG magnetoencephalography, PET positron emission tomography, 2DG 2-deoxyglucose postmortem histology (Reproduced with permission from Grinvald and Hildesheim 2004)

Indeed, the enumeration of salient sites of activity is not necessarily a sufficient level of description of brain function, as significant brief and sparse events may not necessarily show up. Conceptually, this spatial mapping approach was in a large degree inherited from the pre-imaging era, where local lesions, traumas, or tumors provided the main means to chart brain function. Since the mapping approach has partly resulted from the inherent limitations of temporal resolution due to the slow hemodynamic response measured with fMRI, its popularity testifies to how instrumentation, utilization, and conceptualization go hand in hand.

Conversely, electrophysiological methods such as MEG and EEG have steered neuroscience studies away from spatial mapping into an approach where brain function is conceptualized as rapidly evolving distributed spatiotemporal activation patterns. Both MEG and EEG are measures of ongoing neuronal activity and are ultimately generated by the same sources: postsynaptic currents in groups of neurons which have a geometrical arrangement favoring currents with a uniform direction across nearby neurons. The most significant such assembly is that of pyramidal cells in the cerebral cortex. The macroscopic source current generated by these assemblies, often called the primary current (Hämäläinen et al. 1993; Hari and Ilmoniemi 1986), creates an electric potential distribution, which can be sampled on the scalp using EEG.

This potential distribution is associated with passive volume currents in the conducting medium (Hari and Ilmoniemi 1986). In general, the primary and volume currents together generate the magnetic field, measured with MEG. Rather surprisingly, however, the effect of the volume currents can be often quite easily taken into account. The integral effect of all the currents to the magnetic field can be computed accurately with relatively undetailed model of electrical conductivity distribution (Hämäläinen and Sarvas 1989; Okada et al. 1999), whereas EEG is significantly affected by the conductivity details between the sources and the electrodes. Furthermore, the effect of the real or virtual reference electrode employed has to be correctly taken into account. Since MEG and EEG capture electrical activity patterns of neural populations directly, they allow for functional brain activity to be delineated at a very fine temporal scale and may be decomposed into its oscillatory frequency components.

The ability to compute the MEG/EEG patterns generated by known sources, commonly called the solution of the forward problem, opens up the possibility to find an estimate of the primary currents given the MEG measurement. However, this inverse problem is ill-posed: many different current distributions are capable of explaining the data (non-uniqueness), and the solutions are sensitive to noise in the data (the problem is ill-conditioned). In turn, this means that when one wants to go from the MEG recordings at the sensor level to a plausible estimate of its underlying source, one must accept that one has to always employ simplifying assumptions and approximations. If one understands the qualities of this simplified equivalent source description, it is possible to gain useful insights to brain function from it even though the actual complex spatial details cannot be reliably recovered because the measurements are necessarily made far away from the sources. Recent advances in MEG sensor technology discussed in Sect. 4 may help partly overcome this problem and give access to finer spatial detail by bringing the sensors closer to the sources. Incidentally, other noninvasive brain imaging methods have their challenges as well, but the MEG (and EEG) communities have historically been very vocal about expressing their limitations in the literature.

One can mitigate the non-uniqueness of the inverse problem by imposing anatomically and physiologically meaningful constraints. The noise sensitivity can be reduced by regularization: exact match between the measured data and those predicted by model is partly sacrificed to make the estimates more robust (Baillet et al. 2001; Hämäläinen et al. 1993). Interestingly, MEG analysis has from the beginning emphasized the need to work in terms of the source estimates, in the source space, rather than with the sensor-space signals. In contrast, even today, the vast majority of EEG studies rely on traditional sensor-space analyses. The source space approach is, in fact, more straightforward with MEG than with EEG because of the availability of a reasonably accurate forward model and because, as will be discussed below, MEG sees a specific subset of the sources in the brain. One contribution from MEG to the field of neuroscience is hence that the method made it possible to discern the functional events of the brain at a new and anatomically and temporally relevant level of detail that allowed pertinent neuroscientific questions to be pursued. The source estimation approach gradually made its way to EEG analyses as well, emphasizing the benefits of understanding the data in terms of brain sources rather than their manifestations on the scalp or outside. We will discuss the development of source modeling techniques in the following section.

MEG and EEG not only have a high temporal resolution but are measures that reflect the neural currents without the hemodynamic response as a proxy as in the case of fMRI. They are thus able to record electrophysiological activity directly including, for example, the oscillatory characteristics (Adrian 1944; Berger 1929), which are important both in portraying individual responses and in identifying interactions between different sites in the brain. Oscillatory frequencies can thus help distinguish and characterize neural events, e.g., event-related desynchronization and synchronization (ERD and ERS). In addition, oscillations can be also linked together across frequency bands, i.e., the abundance of high-frequency signals may be related to the phase of lower-frequency signals through phase-amplitude coupling. For example, Canolty et al. (2006) used subdural electrodes in humans to show that high gamma amplitude was coupled with theta phase under a variety of tasks and that the degree of coupling depends both on the task and the location where it is measured. This linkage has been subsequently observed noninvasively and has sometimes been used as a measure of local connectivity (Khan et al. 2013) (Fig. 2), which cannot be assessed otherwise due to lack of spatial resolution. Thus, one can sometimes transcend the limits of spatial resolution by employing temporal as well as frequency-domain fingerprints.
Fig. 2

Local connectivity changes in the fusiform face area (FFA). (a) Phase-amplitude coupling (PAC) changes in typically developing (TD) and autism spectrum disorder (ASD) subjects after presentation of different types of stimuli. (b) PAC for emotional faces normalized by PAC for houses (Z-PAC). (c) Z-PAC as in B but between alpha and high gamma only, computed over the entire cortex, for each group. The functionally determined FFA is outlined in bold. This figure demonstrates that even local cortical connectivity changes can be detected with MEG, provided that appropriate measures capitalizing on the spectral and temporal characteristics of the signals are employed

However, it is still extremely difficult to interpret the underlying cellular and circuit level generators of these “macroscale” signals without simultaneous invasive recordings. This difficulty limits the translation of MEG/EEG findings into novel principles of neural information processing or into new treatment possibilities for neurological and psychiatric disorders. There is thus a need to bridge the “macroscale” with the underlying “mesoscale” cellular and circuit level generators. This problem is ideal for biophysical neural modeling where we can have specificity at both scales. To fill this gap, we have employed models which incorporate detailed anatomical and biophysical constraints to generate hypotheses as to the neural origin of observed neocortical brain signals (Jones et al. 2007, 2009; Kerr et al. 2011; Sacchet et al. 2015; Wan et al. 2011). These models also have the utility of linking macroscopic measurements in humans to invasive measurements in animal models and thus make hypotheses about the underlying neural events testable (Sherman et al. 2016).

In limited instances, the phenomena observed noninvasively in MEG or EEG can be also accessed with invasive recordings in humans. In particular, diagnosis of abnormal epileptic signals often requires the use of invasive electrodes either on the surface of the cortex (electrocorticogram, ECoG) or in the brain with isolated depth electrodes or arrays thereof.

An important concept related to noninvasive and invasive measurements is the detectability or signal strength, particularly as a function of the distance between the sensors and the active sources. This depends on three factors: first, the spatial characteristics of the source; second, temporal synchrony within the source area; and third, the spatial selectivity of the sensors. It is well known that at the level of even single neurons or small assemblies of neurons, the currents may exhibit a “closed” or “open” configuration (Lopes da Silva and Van Rotterdam 1992). In addition, the contribution of action potentials as opposed to postsynaptic currents is minor at a distance because of the two opposing current dipoles needed to represent the primary currents corresponding to action potentials (Hämäläinen and Hari 2002). At a macroscopic level, there may be further cancellation effects caused by the macroscopic curvature of the cortex in which the primary currents are normal to the cortical mantle (Ahlfors et al. 2010). In addition to these spatial effects, the strength of the measured signals depends on the length scale on which the activity is synchronous across cells. If this scale is small, measurements made at a distance, effectively averaging the source currents over a larger area or volume, may have a smaller amplitude than the distance dependence determined by the Maxwell’s equations governing the physics would alone predict. This characteristic may explain why high-frequency oscillations, e.g., in the gamma band, are often more easily seen in intracranial measures than in extracranial MEG/EEG (Dalal et al. 2009). Finally, the sensor configuration also needs to be taken into account. This is often described with help of the lead field, which is the sensitivity pattern of the sensor (Tripp 1983). It is worth pointing out that invasive recordings with macroscopic electrodes are in many ways similar to the surface EEG measurements except that some of the electrodes are potentially located much closer to the sources than any electrode on the scalp. In principle, source estimation methods similar to those used in MEG/EEG (Baillet et al. 2001) can be used to estimate the sources of the invasive recordings pending an accurate forward model (Kakisaka et al. 2012; Murakami et al. 2016). In some cases, simultaneous MEG and invasive recordings are also possible, and a combined source estimation approach would then allow better resolution of the brain activity than possible on the basis of the surface measures alone.

Another approach to enhance the spatial information provided by MEG is to try to determine the extent of the sources. Due to the ambiguity of the electromagnetic inverse problem, this is a very difficult task. In this regard, assuming that the sources of MEG will be confined to the cortical mantle can be useful. For example, it might not be possible to explain the observed signals well with a focal source instead of an extended one because the focal source (current dipole) best explaining the data would not be located in the allowed cortical source space. Conversely, if a cortical constraint is not employed, a best-fitting dipole source located in the gray matter very likely indicates that the true activity has a limited spatial extent. It has also been argued (Murakami and Okada 2015) that the current density supported by brain tissue is remarkably constant across species and brain structures: there appears to be a maximum value across the brain structures and species (q0 = 1–2 nAm/mm2). The empirical values presented in Murakami and Okada (2015) closely matched the theoretical values obtained with an independently validated neural network model, indicating that the apparent invariance is not coincidental. This maximum value leads to a lower limit for the source extent. Since the current-dipole density q, (average) dipole amplitude Q, and activated area A are related by q = Q/A and q < q0, we have A > Q/q0. For an estimated current-dipole amplitude Q = 20 nAm, the corresponding active cortical area should be A ≥ 10 mm2. Similar conclusions can be drawn from the EEG source estimates. However, the dependence of EEG on the tissue conductivities, which are not precisely known, makes it difficult to arrive at reliable quantitative estimates of the source strengths.

A third approach to enhance the understanding of the spatiotemporal characteristics of brain function is to combine electrophysiological measurements with fMRI. Even though fMRI relates to the electrical activity measured by MEG and EEG indirectly due to the hemodynamic coupling, it is attractive to try to benefit from the high spatial resolution of fMRI in MEG and EEG studies by combining the two types of measures in the analysis. In such an approach, several strategies are possible (Horwitz and Poeppel 2002): (i) comparison to obtain converging evidence, (ii) direct data fusion, and (iii) computational neural modeling using a generative model capable of predicting both fMRI and MEG/EEG given the sources.

Up to recently, most studies have been limited to the first two alternatives. While comparison is at least in principle straightforward, data fusion must take into account the possibility of missing sources in either electrophysiological or hemodynamic measures. Such missing sources may be directly due to electrophysiological characteristics of the sources or even biophysics underlying the two measurement techniques.

The earliest data fusion model used fMRI as an a priori weight in the computation of cortically constrained source estimates (Dale et al. 2000; Liu et al. 1998). In the application of this method, the fMRI constraint was usually based on a generous “omnibus” contrast to include all possible source candidates, and sources outside the fMRI-defined areas were allowed by making the weighting only partial (Liu et al. 1998). While more complicated approaches have since been introduced, e.g., Daunizeau et al. (2007a, b) and Ou et al. (2010), they have not made their way to standard MEG/EEG analysis streams due to computational complexity and challenges in data acquisition. In some cases, the comparison yielding even non-converging evidence may be more useful because the discordant information may give important insights due to the different sensitivities of fMRI and MEG/EEG to different types of brain activity (Agam et al. 2011). In addition, in one instance (Agam et al. 2011), another study found evidence concordant with MEG/EEG from direct recordings in monkeys (Heilbronner and Platt 2013), thus underlining the importance of including the comparison method at least as an alternative to data fusion.

In addition to the three classic approaches cited above (Horwitz and Poeppel 2002), it is also possible to link the electrophysiological and hemodynamic measures without knowledge of even a simplified form of a generative model or hemodynamic coupling. Using a combination of machine learning and representational similarity analysis(RSA), fMRI and MEG/EEG can be linked using the representational characteristics generated by experimental stimuli or events (Cichy et al. 2014, 2016). By abstracting away from each domain’s particular source space (e.g., sensor activity patterns in MEG, BOLD responses in fMRI, or behavioral responses) into a common similarity space that is defined by between-condition dissimilarities in activity patterns and response patterns, RSA compares the geometry of the between-condition similarities along the similarity dimension (Kriegeskorte and Kievit 2013; Kriegeskorte et al. 2008).

3 Estimating the Current Sources Underlying MEG

As discussed above, the intricate time-frequency characteristics of MEG and associations (connectivity) between regions of activity, together with the ill-posed nature of the inverse problem, make it difficult to prescribe a universally applicable approach to MEG analysis. Indeed, the spectrum of available MEG analysis methods may be overwhelming or even frustrating to newcomers. However, the existence of many different approaches may also be considered an important benefit. Once the experiment has been established with tentative hypothesis, one can outline the optimal analysis pipeline. It is also conceivable that the experimental design can be informed by the available analysis approaches. Such iterative interaction clearly requires the existence of interdisciplinary research teams, which have been one of the common characteristics of many leading MEG centers up to date.

3.1 Forward Models

The central concept in the MEG analysis is source estimation or inverse modeling, already discussed in general terms above. For the inverse modeling task to succeed, we need to possess an accurate enough forward model, which consists of a description of the distribution electrical conductivity of the head and an analytical or numerical method to compute MEG (and EEG) given the conductivity assumptions and the elementary sources, the source model. Notably, the magnetic permeability of the head is close enough to that of the vacuum, and the time derivatives can be ignored from the Maxwell’s equations for the computation of MEG. Therefore, the quasistatic approximation (Hämäläinen et al. 1993; Plonsey 1969) is sufficient. For the following discussion, it must be noted that the electrical conductivity is not only a location-dependent scalar quantity but may possess anisotropy: the conductivity is dependent on the direction it is measured. This anisotropy can be measured not only with direct electrical means but also with help of diffusion-weighted MRI (Tuch et al. 2001).

The simplest and a surprisingly good approximation for the head’s conductivity distribution for MEG relies on spherical symmetry: the head is assumed to be composed of spherical shells with different electrical conductivities. In this case, a closed-form analytical expression exists for the magnetic field (MEG) outside the sphere (Cuffin and Cohen 1977; Ilmoniemi et al. 1985; Sarvas 1987). This sphere model yields MEG characteristics of remarkable simplicity. First, only currents tangential to the surface contribute to the magnetic field. Second, the radial component of the magnetic field can be computed directly from the primary current, and even the tangential components can be computed without explicit reference to the volume currents. Third, unlike for EEG, the result is independent of the conductivity profile along the model’s radius. While these assertions are not strictly correct under more realistic conductivity assumptions, they serve as a good baseline with respect to which differences can be often considered as small perturbations. Furthermore, while the analytical approach is by far the most efficient and accurate one for the sphere model, the numerical methods described below can be applied in the spherically symmetric case as well, emphasizing the dichotomy between the model assumptions and the actual solution method used. It should be also noted that other simple conductor shapes, including the ellipsoid, can be handled analytically (Cuffin and Cohen 1977).

If the spherical symmetry is abandoned and the head is assumed to consist of homogeneous compartments of realistic shape, the solution can be obtained with the boundary element method(BEM) (Hämäläinen and Sarvas 1989; Mosher et al. 1999). This numerical approach has been finessed and can now even take into account thin layers of CSF (Stenroos and Nummenmaa 2016). In addition, it has been shown that BEM can incorporate compartment topologies earlier believed to be accessible with finite element and finite difference methods (FEM and FDM) only (Stenroos 2016). Specifically, the limitation to nested (layered) and island-in-the-sea geometries can be relaxed. The latter development is particularly important for the infant head: the infant skull consists of separate pieces that are connected by soft tissue (fontanels). In this geometry, each piece of skull shares its single closed boundary surface with the scalp, fontanel, and brain, resulting in piecewise constant conductivity contrast across this boundary. Previous attempts to model this geometry using BEM have employed an approximation in which the fontanel is taken into account with a thinner region in the skull (Roche-Labarbe et al. 2008). For the purposes of MEG modeling in adults, it is, however, often sufficient to consider the skull to be a perfect insulator (Hämäläinen and Sarvas 1987, 1989). However, it has been later established that a three-compartment model consisting of the intracranial space, the skull, and the scalp is preferable even for MEG (Stenroos et al. 2014) and certainly if combined analysis of MEG and EEG is contemplated. Routine use of multi-compartment models is, however, dependent upon using accurate and reliable MRI-based means to estimate the shape of the skull compartment. Furthermore, the determination of the electrical conductivity of each compartment remains a challenge.

The prevailing approach in computing the MEG/EEG forward solutions in complex conductor geometries is the finite element method(FEM), which has been developed to a great sophistication (Drechsler et al. 2009; Lanfer et al. 2007; Lew et al. 2009, 2013;Wolters et al. 2007a,b,c). The FEM can incorporate an arbitrary conductivity distribution, including anisotropies. However, especially given the uncertainties in the conductor geometry, the actual electrical conductivities, and possibly the need to resort to atlas-based approximate models, the modern BEM approaches offer several benefits: (i) Since the potential is computed on the boundary surfaces only, the intricate methods to accommodate the source singularities in the volume-based FEM are not needed. (ii) Thin layers of CSF or touching surfaces can be accommodated by locally increasing the density of the surface tessellations rather than having to create a large number of voxel elements. (iii) The computational burden is small enough to allow a detailed study of the effects of the conductivity geometry, conductivities, and surface tessellation density on the accuracy of the solution. On the other hand, the FEM offers the capability to model anisotropic conductivity, possibly important in the white matter (Gullmar et al. 2010) and the skull (Dannhauer et al. 2011).

The refinement of the forward model increases the accuracy of the source locations estimated (reduces bias). However, in general an improved forward model does not increase the spatial resolution, if understood as the ability to resolve two close by simultaneously active sources. In some cases, the source estimates may actually be relatively insensitive to the accuracy of the forward model (Stenroos and Hauk 2013). Furthermore, as mentioned earlier, the forward model embodies the biophysics of the MEG signal generation and is crucial for understanding the spatial characteristics of MEG. In other words, the forward model is an expression of fundamental laws of physics and is thus independent of the neuroscience per se or the clinical question being addressed. One can actually gain a lot of useful insight to the characteristics of MEG (and EEG) by studying the solutions of the forward problem. For example, one can study the relative sensitivities of MEG/EEG to sources at different cortical sites and gain understanding to the relative merits of the two types of measurements with help of just the solution of the forward problem without a need to specify or apply a source estimation procedure first (Goldenholz et al. 2009).

3.2 Source Estimation

In general, every MEG sensor sees every source in the brain with different weights, and thus the signal seen in any MEG sensor is a linear combination of the time courses of all sources. The aim of the solution of the inverse problem is to produce source estimates, which correctly describe the locations and extents of the sources underlying the measured MEG data, and yield the unmixed waveforms of the underlying sources. This task is complicated by the non-uniqueness discussed in the previous section: there exist silent sources, invisible in MEG, EEG, or both. The MEG/EEG source estimation methods can be divided into three categories: (i) parametric source models, (ii) distributed current estimates, and (iii) scanning approaches.

In the parametric approach, one commonly assumes that the cortical activity underlying the measurements is sparse, i.e., salient activity occurs only in a small number of cortical sites, and that each active area has a small enough spatial extent to be equivalently accounted for by a point source, a current dipole. This multidipole model has been very successful in the analysis of evoked potentials and fields. Even though the multidipole model is often used to explain measurements of primary and secondary sensory responses, it has also been employed in modeling of more complex cognitive functions (see, e.g., Nishitani et al. 2004 and Salmelin et al. 1994) and cortical rhythms (Salmelin and Hari 1994).

The distributed modeling approaches assume a distribution of sources on the cortex and other structures and apply an additional criterion to select a particular distribution to explain the data and to produce an image of the most likely current distribution. To date, the most successful approach of this kind is the cortically constrained minimum-norm estimate (MNE) (Dale et al. 2000; Dale and Sereno 1993; Hämäläinen and Ilmoniemi 1984), which constrains the currents to the cortical mantle and selects a solution, which has the minimum overall power. The MNE is a diffuse estimate, usually overestimating the extent of the source, and, therefore, the extent of the MNE should not be interpreted literally. In addition to the L2-norm constraint employed in MNE, it is possible to use a sparsifying criterion, e.g., the L1-norm to produce solutions which resemble multiple dipole models with the difference that the constellation of sources is different at each time point (Uutela et al. 1999). Subsequent developments of L1-norm solutions have included constraints on the source waveforms and requirement that the set of sources remains unchanged throughout the analysis period (Gramfort et al. 2012, 2013; Ou et al. 2009).

In the third approach to source estimation, a suitable scanning function, derived from the input data, is evaluated at each candidate source location. A high value indicates a likely location of a source. Examples of this method are the linearly constrained minimum variance beamformer (Sekihara and Nagarajan 2008; Van Veen and Buckley 1988) and MUSIC (Mosher and Leahy 1998; Mosher et al. 1992) approaches. The beamformer method has gained a lot of popularity among MEG researchers, while its use in EEG has been limited. This is likely due to the fact that for the beamformer method to work, the forward model needs to be sufficiently accurate (Steinstrater et al. 2010). Finally, the scanning approaches differ from the parametric dipole model and the source imaging approaches in the sense that the “pseudo-images” they produce do not constitute a current distribution which is capable of directly explaining the measured data.

During the past 30 years, all three types of methods have been widely used for the analysis of cortical activity and have also been validated to varying degrees in patients with invasive recordings (see, e.g., Tanaka et al. 2010). However, subcortical structures and the cerebellum also play important roles in brain function. For example, brainstem and thalamic relay nuclei have a central role in sensory processing (Jones 1998, 2001). Thalamocortical and hippocampal oscillations govern states of sleep, arousal, and anesthesia (Steriade et al. 1993). Striatal regions are crucial for movement planning, while limbic structures like the hippocampus and amygdala drive memory, emotion, and learning (Alexander et al. 1986; Graybiel 2000; Phelps and LeDoux 2005). Unfortunately, the anatomy of the brain poses two particular challenges for deep source estimation with MEG. First, deep brain structures are farther away from the sensors than the cerebral cortex and thus produce much lower-amplitude MEG signals than the cortex. A second, perhaps more fundamental problem stems from the fact that the subcortical structures are surrounded by the cortical mantle. As a result, measurements arising from the activity of deep structures can, in principle, be explained by a surrogate distribution of currents on the cortical surface. This ambiguity also means that it is even harder to estimate subcortical activity when cortical activity is occurring simultaneously.

However, in our recent paper (Krishnaswamy et al. 2017), we reason that these limitations could be mitigated if only a finite number of cortical sites are active together with the subcortical structures. In many neuroscience studies, salient cortical activity at any moment in time tends to be restricted to a small set of well-circumscribed areas. It follows that if we can identify this sparse subset of active cortical sources and eliminate the remaining irrelevant cortical sources (Babadi et al. 2014), we have a chance at recovering the locations and time courses of both cortical and subcortical sources. In Krishnaswamy et al. (2017), we demonstrate the feasibility of this approach by introducing a new source estimation method capitalizing on this insight. Its general applicability will depend upon the degree of overlap among the MEG field patterns of these candidate sources and the signal-to-noise ratio of the measurements.

4 MEG Instruments

4.1 Introduction

In his 1944 paper entitled Brain Rhythms (Adrian 1944), Edgar Adrian states:

That is a very long way from saying that the electroencephalogram can tell us how the subject will think and act. In fact the information which it gives relates to a very limited field. But the limitation arises mainly from the fact that we can only record the gross effects and not the detailed patterns in the brain. With present methods the skull and the scalp are too much in the way, and we need some new physical method to read through them. … In these days we may look with some confidence to the physicists to produce such an instrument, for it is just the sort of thing they can do.

As we have seen in the previous sections, MEG is an electrophysiological recording method, which can, indeed, record brain activity with little influence by the intervening tissues. In particular, detailed information about the conductivity distribution in the head is generally not needed, and for many purposes, the skull can be even approximated by a perfect insulator and the scalp can be omitted from the model. The invention and development of the ultrasensitive SQUID sensors enabled real-time MEG recordings and the building of multichannel MEG instruments. In this section, we will discuss recent developments in MEG instrumentation, including innovations in the use of regular SQUID sensors operating in liquid-helium temperature (4.2 K, low-Tc SQUIDS) as well as high-Tc-SQUIDs operating at higher (liquid nitrogen) temperatures.

In addition, totally different sensor technologies have been proposed and demonstrated. These include atomic vapor cell (optically pumped) magnetometers, which employ a hot gas of alkali metal atoms to sense magnetic fields. The most sensitive flavor of such devices, spin-exchange relaxation-free (SERF) magnetometers, can demonstrate sensitivities of approximately \( 0.5\ \mathrm{aT}/\sqrt{\mathrm{Hz}} \). (Kominis et al. 2003). In 2012, a team of National Institute of Science and Technology (NIST) investigators demonstrated a single-channel chip-scale atomic magnetometer capable of detecting magnetic fields generated by an intact human brain (Sander et al. 2012). The atomic vapor magnetometer approach to MEG continues to be pursued by various groups around the world (Kim et al. 2014). Magnetometers based on the nitrogen vacancy (NV) defects in diamond are a relatively recent but fast evolving technology. First proposed in Taylor et al. (2008), such magnetometers have since surpassed all other magnetometers except for SQUIDs and vapor cell magnetometers in sensitivity and can currently demonstrate sensitives on the order of \( 1\ \mathrm{pT}/\sqrt{\mathrm{Hz}} \). A recent paper reported successful detection of the magnetic field of a single firing neuron from outside an intact marine worm (Barry et al. 2016). Theoretical calculations show that the sensitivities high enough for MEG measurements will be possible with the NV technology, but the feasibility has not yet been demonstrated experimentally.

It is indeed of utmost importance that the MEG instrumentation develops in parallel with our understanding of the brain’s functional and structural organization. With sensors that capture details beyond the present spatial limits, neuroscientists may discern yet undetected events and disentangle known events using better resolution available in time, frequency, and space. Below, we will consider the feasibility of using these newer sensor technologies in light of recent demonstrations and the benefits from the ability of bringing the sensors to the close proximity of the scalp.

4.2 Pushing the Limits of Conventional MEG Systems Based on Low-Tc SQUIDs

While the first practical MEG devices employing low-Tc SQUID sensors operating at liquid-helium temperatures were introduced more than 45 years ago, innovations improving the spatial resolution, sensitivity, and usability of systems based on these sensors continue. Among all presently available magnetic field sensors that can be potentially used for MEG, the low-Tc SQUIDs have unique benefits: (i) the sensitivity, especially in the low frequency range below a few Hz, is exquisite; (ii) once operational, a MEG system employing low-Tc SQUIDs is very stable and reliable; and (iii) the manufacturing techniques of low-Tc SQUIDs are well established, and the produced sensors are reliable and of uniform quality. Some of the challenges of the helium-temperature technology are the need for a dewar, which results in a large distance (at least 20 mm) between the sensors and the brain, and the actual need for helium refills, which increases the cost of the operation of such a system. Despite the emergence of several hardware- and software-based noise cancellation approaches, best quality data are still obtained in a heavily shielded expensive magnetically shielded room. Furthermore, the fixed size of the dewar and the sensor helmet makes the adult MEG systems suboptimal for the study of younger subjects with smaller heads.

For MEG recordings from newborns and infants, many of the drawbacks of MEG systems based on low-Tc SQUIDs have been mitigated in the 375-channel, whole-head MEG system (“BabyMEG”) we recently developed for studying the electrophysiological development of human brain during the first years of life (Okada et al. 2016). The helmet accommodates heads up to 95% of 36-month-old boys in the USA. The unique two-layer sensor array consists of 270 magnetometers (15 mm coil-to-coil spacing) in the inner layer and 35 three-axis magnetometers in the outer layer, 4 cm away from the inner layer. In addition, there are three three-axis reference magnetometers about 25 cm away from the primary sensor array. With the help of a remotely operated position adjustment mechanism, the sensor array can be positioned to provide a uniform short spacing (mean 8.5 mm) between the sensor array and room-temperature surface of the dewar. A closed-cycle helium recycler (Wang et al. 2016) provides a maintenance-free continuous operation, eliminating the need for helium, with no interruption needed during MEG measurements. Ongoing spontaneous brain activity can be monitored in real time without interference from external magnetic noise sources including the recycler, using a combination of a lightly shielded two-layer magnetically shielded room, an external active shielding, a signal-space projection method, and a synthetic gradiometer approach. Evoked responses in the cortex can be clearly detected without averaging.

This system demonstrates that with conventional low-Tc SQUID technology, it is possible (i) to eliminate the need for helium refills with a continuously operating helium recycler, (ii) to construct a system employing simple magnetometers only and operate it in a lightly shielded environment, (iii) to decrease the distance from the detector coils to outside of the dewar significantly, and (iv) with advanced noise cancellation techniques to reduce the noise to a level where spontaneous brain activity and single-trial evoked responses can be studied. Furthermore, the system is specifically tailored for the infant population with a small helmet and a high detector density. It can be envisioned that an analogous system could be produced for the adult population with significant performance gains, thanks to the short standoff of the detectors from the scalp. It can be estimated that the smaller detector-to-brain distance in such a system would justify increasing the number of channels from the present 250–300 to about 500. We will next consider the opportunities for developing even better MEG systems using other sensor technologies.

4.3 The Promise of On-Scalp MEG

As discussed above, the sensitivity and reliability of low-Tc SQUID detectors is still unsurpassed. However, there are two remaining challenges in their use as MEG sensors. The first is the relatively long distance between the sensors and the scalp (typically 20–40 mm (Andersen et al.)). Related to this, a second challenge is the fixed size of the MEG helmet, which means that for young subjects and even for adults with smaller heads or atypical head shapes, the distance from the sensors is even larger. There have been attempts to construct MEG systems with multiple sensor units, but because of the large size of the liquid-helium dewars, these devices have turned out to be difficult to operate. Therefore, small “magnetrode” sensors or small sensor arrays which could be positioned in a customized sensor holder would be highly desirable. With such devices, an on-scalp MEG(OSMEG) system can be constructed with a helmet optimized for each individual. Possible sensor technologies for OSMEG include high-Tc SQUIDs, optically pumped magnetometers (OPMs), and possibly in the future the NV diamond sensors.

A recent publication (Iivanainen et al. 2017) reported simulations to predict the performance of an OSMEG array constructed with OPMs. The capabilities of an OPM array were compared to those of a conventional commercial 306-channel low-Tc instrument. Three types of OPM arrays were considered: one measuring the normal field component (nOPM), another two tangential field components (tOPM), and a third (aOPM) measuring all three components of the magnetic field at each detector site. The closest point of each OPM sensor was 1 mm away from the scalp, while the sensors of the low-Tc system were ≥ 20 mm away. The nOPM array had 102, the tOPM array had 204, and the aOPM array had 306 sensors. Therefore, the sensor density was somewhat higher in the aOPM array than in the conventional low-Tc array due to the fact that the area on which the OPM sensors were placed was smaller. The noise levels were assumed to be \( 3\ \mathrm{fT}/\sqrt{\mathrm{Hz}} \), \( 3\ \mathrm{fT}/\mathrm{cm}\sqrt{\mathrm{Hz}} \), and \( 6\ \mathrm{fT}/\sqrt{\mathrm{Hz}} \) for SQUID magnetometers, planar gradiometers, and OPMs, respectively. Ten adult heads were considered, and the sensitivity to cortical currents and the capability to estimate sources on the cortex was assessed. The signal power was found to be 5.3 to 7.5 times higher in the OPM array than in the SQUID array, while the correlations between field patterns of dipoles were reduced by a factor of 2.8 to 3.6. The information-theoretical channel capacities (Kemppainen and Ilmoniemi 1989) of the OPM arrays were clearly higher than those of the SQUID array. The dipole localization accuracies of the arrays were similar, while the point spread of minimum-norm estimates was about 2.5 times higher in the SQUID array than in the OPM arrays.

We also recently performed similar simulations to compare the performance of high-density on-scalp MEG (“OSMEG”) and EEG (“HD-EEG”) arrays to that of the same commercial system (Elekta-Neuromag “VectorView”) as in Iivanainen et al. (2017). The arrangement of the three detector arrays is shown in Fig. 3. Using the standard three-compartment boundary element forward model (Hämäläinen and Sarvas 1989) and a dense set of current dipoles normal to the cortical mantle in an adult brain (Ahveninen et al. 2006), we constructed the gain matrices for OSMEG, VectorView, and HD-EEG arrays. We first evaluated the singular value profiles of the three gain matrices. The number of singular values higher than 0.01 times the largest one was 182 for OSMEG, 109 for VectorView, and 81 for HD-EEG. At 0.001 cutoff, the corresponding values were 387, 240, and 188. Consistent with Iivanainen et al. (2017), this result clearly indicates that on-scalp MEG carries a superior amount of spatial information with respect to the other two arrays considered.
Fig. 3

(Left) Possible sensor layout for a 500-channel on-scalp MEG system. (Middle) The locations of the sensors in a commercial 306-channel MEG system (Elekta-Neuromag Vectorview). (Right) Locations of evenly distributed 500 EEG electrodes on the scalp

We then proceeded to take into account the effect of noise. We assumed a 1–40-Hz bandwidth, and on the basis of VectorView empty room noise in our well-shielded facility, we employed \( 5\ \mathrm{fT}/\sqrt{\mathrm{Hz}} \) for VectorView magnetometers, \( 3.9\ \mathrm{fT}/\sqrt{\mathrm{Hz}} \) for the planar gradiometers in the same system, and \( 8.3\ \mathrm{fT}/\sqrt{\mathrm{Hz}} \) for the OSMEG. A noise STD of 2 μV was assumed for the EEG measurements corresponding to a noise density of \( 320\ \mathrm{nV}/\sqrt{\mathrm{Hz}} \). Figure 4 shows localization errors and overall SNRs for three cortical patches with varying cortical surface current densities. It is clear that the OSMEG system outperformed the other two in all cortical areas considered.
Fig. 4

Localization errors (left) and SNRs for the three arrays shown in Fig. 3. The actual source was a patch of cortex in the frontal lobe (top), in the somatosensory hand area (middle), and in the visual cortex (bottom). The cortical current density was parametrically varied; unaveraged raw data with 40-Hz bandwidth was assumed

To further elucidate the overall performance, we also conducted a Monte Carlo simulation with current dipoles situated at each vertex of the cortical tessellation, oriented normal to the cortical mantle, and noise corresponding to the above estimates added. We repeated the dipole localization 50 times with different noise realizations and computed the localization errors as a function of dipole location (see Fig. 5). Remarkably, even the VectorView array outperformed the HD-EEG for superficial sources, and the OSMEG array was clearly superior with respect to the other two. For deeper sources, EEG provided some gain with respect to MEG. At the crests of the gyri and at the troughs of the sulci, there were larger localization errors in MEG than in EEG, but this “invisible” part of the cortex was minimized by the use of the OSMEG array, as expected.
Fig. 5

Dipole localization errors for the three arrays, shown in Fig. 3, considered as a function of the source location on the cortex. The data are visualized on an inflated view of the lateral and medial surfaces of the left hemisphere. The strength of each dipole in the forward simulation was 10 nAm, and 40 averages in an event-related study with a 40-Hz bandwidth were assumed

Both of the simulations presented clearly demonstrate the substantial gains achievable with OSMEG. The signal-to-noise and the information-theoretical channel capacity are increased. Related to the latter, the number of large singular values of the gain matrix is increased, and as a result, the field patterns of sources close to each other can be discerned. Furthermore, all performance indices related to source estimates are improved. Some researchers have even nicknamed OSMEG as “noninvasive ECoG,” since spatial resolution comparable to this invasive method can be achieved. Notably, even a high-density EEG array does not seem to be able to outperform OSMEG because the EEG electrodes are already positioned on the scalp, i.e., as close as possible to the sources of the signals. The remaining main challenge in EEG for higher spatial resolution is the smearing of the potential patterns due to the layered structure of the intervening tissues (brain-skull-scalp), which cannot be circumvented since it is a physical fact. Finally, the simulations have not yet considered an additional potential benefit of OSMEG: the lead fields on-scalp sensors will be more focused on the cortex than those of conventional MEG or EEG. If the cortical activity, especially at higher frequencies, is coherent only over small distances, there is a strong temporal cancellation of signals due to incoherent sources within the wide high-sensitivity region of the distant MEG sensors, and as a result, the SNR suffers. We may thus expect that OSMEG may not only be able to detect weaker signals and locate their sources more accurately but also give better sensitivity to high-frequency activity.

4.4 Experimental Demonstrations and Outlook

The opportunities for higher-quality noninvasive recording of MEG signals have also been demonstrated with experiments. These demonstrations include measurements with both high-Tc SQUIDs (Andersen et al. 2017; Faley et al. 2012; Xie et al. 2017) and OPMs (Borna et al. 2017; Boto et al. 2017). Most recently (Boto et al. 2018), it has been demonstrated that with OPMs, it is possible to construct a practical “wearable” MEG system, in which the sensors are fixed to the head and thus compensation for head movements is not necessary. However, the latter requires compensation for both the homogeneous magnetic fields and field gradients using a specially designed compensation coil system (Holmes et al. 2018). Since the noise performance of OPM sensors, which detect the absolute magnetic fields, deteriorates in a dc field, the homogeneous field compensation system is in general required even for measurements where the sensor array is stationary.

None of these early demonstrations of the feasibility of new MEG sensor technologies have revealed fundamentally new information about brain activity. This is at least partly due to the relatively high noise level of the sensors (≥\( 15\ \mathrm{fT}/\sqrt{\mathrm{Hz}} \)), which partly masks the benefits due to high spatial frequencies present in the on-scalp MEG data. Nonetheless, the simpler or nonexisting cryogenics and the capability of using the (OPM) sensors in a customized helmet attached to the head open new exciting avenues for MEG studies and for widespread acceptance of MEG.

Unlike the SQUID magnetometers, the OPMs can tolerate strong transient magnetic fields and recover to normal operation in a short amount of time. This means that it is possible to construct an instrument which records MEG signals subsequent to transcranial magnetic stimulation. Under support from the NIH (National Institute of Neurological Disorders and Stroke, Grant 5R44NS090894–04), we are presently building a combined multichannel TMS (16 channels) and MEG (25 channels) array, which employs OPMs for MEG. This system will enable changing the stimulation site of TMS quickly and allow for the recording the ensuing brain signals.

5 The Usefulness of MEG

In the previous sections, we have highlighted MEG and EEG as unique tools for measuring the spatial, temporal, and spectral characteristics of neuronal brain activity and for elucidating the functional organization of the brain. In retrospect, the emergence of whole-head MEG measurements in the early 1990s is a central milestone in the history of MEG. Until then, MEG was a small and experimental specialist field with very little impact compared with the related electric potential measurements, the EEG. Since the 1990s, however, the history of MEG demonstrates a series of unique contributions in charting new territories (Hari and Salmelin 2012), for instance, by advancing the understanding of brain responses following auditory (Sams and Hari 1991), somatosensory (Hari and Kaukoranta 1985), and visual (Ahlfors et al. 1992; Aine et al. 1996; Brenner et al. 1975; Sharon et al. 2007) stimulation. Such studies have effectively capitalized on the ability of MEG to resolve the spatiotemporal characteristics of events at the source level with good accuracy and reliability and have also inspired new ways for making use of EEG recordings.

Today, MEG can be viewed as a well-established albeit small field, with about 200 active MEG labs globally, which is a small number compared to the approximately 30,000 MRI (some indicate over 40,000) installations and the countless EEG labs. Today, MEG is used in several research domains, such as (i) basic neuroscience aiming at advancing the understanding of how the brain is functionally organized in terms of oscillations, networks, and network relationships; (ii) cognitive neuroscience aiming at increasing the basic understanding of cognitive processes such as perception, attention, memory, language, and executive control; (iii) clinical neuroscience aiming at identifying disease- and severity-related brain function characteristics in clinical cohorts such as Parkinson’s and Alzheimer’s diseases, autism spectrum disorders, schizophrenia, ADHD, epilepsy, and TBI; (iv) clinical applications, utilizing MEG as a tool for presurgical planning of refractory epilepsy and presurgical functional mapping of sensory and language processing and also including emerging applications such as studies of TBI and dementia; and (v) instrumentation research aiming at exploring and benchmarking what new aspects of brain activity new MEG sensor technologies (e.g., on-scalp high-Tc SQUIDS and OPMs) may reveal. For all areas of application, the usefulness of MEG has built primarily on the ability to discriminate between activity patterns at a millisecond temporal resolution, combined with a spatial accuracy in the 0.5–2 cm range and a spatial discriminability in the millimeter range. As discussed above, advances have been enabled by vigorous method development, which includes methods for preprocessing, modeling of MEG signals in terms of their underlying sources, statistical evaluation of the results, and computational modeling of the estimated cerebral activity. There have also been significant advances in instrumentation, including the use of new sensor technologies (high-Tc SQUIDs and optically pumped magnetometers) and development of hybrid systems capable of acquiring both MEG and ultralow-field MRI data (Vesanen et al. 2013) or recording MEG following TMS.

5.1 MEG and EEG in the Literature

Given the small number of active MEG laboratories, it is obvious that MEG is not yet a commonly used method, while EEG has enjoyed widespread use for more than 60 years. However, although MEG is not widely used at present, there are signs that it is gaining ground in terms of influence. In the following section, we outline bibliometric data from PubMed to illustrate the growth of MEG and its influence relative to EEG and also relative to the publications that cite MEG and EEG publications.

5.1.1 Number of Publications

By 2018, PubMed indexes more than 28 million publications in biomedicine and life science and covers the vast majority of MEG and EEG publications. Between 1995 and 2015 (where reliable metrics are available for estimates of citations), the global number of scientific publications in PubMed has grown by 5% annually, from about 500,000 publications per year in 1995 to about 1,100,000 publications per year in 2015. Looking at EEG and MEG publications within the same time frame, we find that EEG follows this general 5% growth trend (from ca. 2000 publications in 1995 to ca. 5000 in 2015), while MEG shows a steady 8% growth (from ca. 100 publications in 1995 to ca. 400 in 2015). MEG is thus by far a smaller field than EEG, but has a steady growth pattern surpassing that of science in general and also that of EEG. Still, the actual volumes of MEG publications are approximately at the level of EEG in the 1950s, and there is a lot of room for growth. However, from the perspective of the individual researcher, the primary bibliometric measure of interest is not the sheer numbers of publications produced by a field, since this mainly reflects the availability of the equipment per se, but rather is the frequency in which publications are cited.

5.1.2 Influence via Citations

A publication can be influential in terms of citations in at least two ways: either by a first-order effect, by the number of publications that cites it, or by a second-order or “ambassador” effect, where one considers the extent to which the citing publications are themselves cited and could thereby channel interest in the publications they cite. By combining PubMed statistics on citations with data from Web of Science, we can look at these two metrics of influence and compare the general influence of the 6409 MEG and the 71,562 EEG publications indexed within the PubMed period and also assess how this has developed over time.

First-Order Influence
As we show in Table 1, for all comparisons of citations between EEG and MEG published between 1995 and 2015 (mean, median, top percentile, or share never cited), the data are systematically in favor of MEG with an average advantage over EEG of about 35%. Not surprisingly, therefore, the annual increase in citations shows numbers in MEG’s favor too. Notably, the annual increases also indicate a possible difference in the lifetime of publications, as the increase in citations seems to plateau for EEG after 10 years so that publications more than 10 years old no longer increase in citations (EEG publications between 1995 and 2005 showing a yearly increase of 0.1 citation, as compared with 3.2 for younger EEG publications), while there is a steady annual increase also for MEG publications 10–20 years old similar to that of 0–10-year-old publications (MEG publications between 1995 and 2005 showing a yearly increase of 2.7 citations and 3.7 for younger MEG publications).
Table 1

First-order influence

Citations

EEG

MEG

MEG advantage

Mean

25.7

32.3

26%

Median

11

14

27%

Share never cited

7%

4.5%

36%

Annual increase 1995–2015

1.9

3

58%

 1995–2005

0.1

2.7

 

 2006–2015

3.2

3.7

 

Top 0.1%

1120

1552

39%

Top 1%

408

550

35%

Top 10%

131

170

30%

Second-Order Influence
By also looking at the extent to which the citing publications are themselves cited, we can then ask questions such as whether there are differences in the “clout” of the publications that cite MEG and EEG publications, respectively. To answer this, we compare the influence of the 57,512 publications that cite the 6409 MEG publications and the 334,285 publications that cite the 71,562 EEG publications indexed in PubMed (all from between 1995 and 2015). As we show in Table 2, there are systematic differences in mean, median, share never cited, and yearly increase to the favor of MEG also here, albeit to a smaller degree, with an average advantage to MEG over EEG of about 20%.
Table 2

Second-order influence

Citations of citations

EEG

MEG

MEG advantage

Mean

26.3

27.9

 6%

Median

9.5

11.9

25%

Share never cited

12%

7%

42%

Yearly increase 1995–2015

2.1

2.3

10%

 1995–2005

1.6

2

 

 2006–2015

2.3

2.5

 

5.1.3 Conceptual Influence

The fate of a publication is of course determined by many different factors in addition to first- and second-order citations, and a significant factor is whether the publication surfaces at all in literature searches that contain search terms such as “magnetoencephalography” or “electroencephalography.” Since all PubMed publications are manually tagged by MeSH (Medical Subject Headings) content tags, from which “magnetoencephalography” and “electroencephalography” are two, we can tease apart the bibliometric data for “MEG” publications that relate to EEG as well, from those that do not. This allows us to understand the influence of publications that have relevance to readers from both the fields of MEG and EEG and from those that are more exclusively relevant to the MEG field. Below, we show some core bibliometric data from such selections for MEG publications between 1995 and 2015, both in terms of first- and second-order influence.

As we show in Table 3, there is an about 25% increase in the mean number of citations for MEG publications that also are MeSH tagged with EEG. Such a tag then means that the publication also includes or discusses EEG in addition to MEG and that it thereby surfaces in literature searches using “MEG” as well as “EEG.” Notably, this wider readership shows the opposite direction of results for second-order effects on citations, so that the publications that cite “MEG only” publications are themselves cited more often (29.4) than those that cite MEG + EEG publications (25.9). This pattern most likely stems from the fact that a MEG publication that concerns EEG also cites EEG publications to a larger extent than MEG only publications: since EEG publications in general show lower first- and second-order citations scores than MEG, an increase in EEG citations to the publication thereby lowers its second-order influence.
Table 3

Conceptual influence

Citations

MEG − EEG

MEG + EEG

MEG + EEG advantage

Citations mean

30.6

38.0

24%

Citations of citations mean

29.4

25.9

−12%

5.2 Conclusions

The total number of publications listed PubMed increases by 5% annually. The corresponding rate for EEG publications is the same, while the number of MEG publications increases at an annual rate of 8%. From the data on first- and second-order influence from publications, we learn that MEG publications have a systematically better momentum and longevity than EEG publications. In addition, MEG publications typically have better momentum than the articles that cite them. Finally, we learn that MEG publications, which also are MeSH tagged with EEG and show up to a broader audience in literature search, have better momentum than those that exclusively concern MEG. This indicates that a MEG publication, and especially if it also concerns EEG, hence has a comparatively strong influence. This may be due to the fact that MEG publications have not only resulted in new neuroscience data but also often include new methodological ideas and advances which have inspired related fields, in particular that of EEG. A recent review on the role of MEG (Baillet 2017) also includes comparisons of the impact of MEG and other imaging methods, including fMRI, thus providing additional insights.

6 How to Run a Successful MEG Study? Words of Advice from Experienced MEG Investigators

While the unique virtues of MEG might be clear and attainable by a seasoned user, a MEG novice may find the method complex and difficult to master. To be successful, all aspects of a MEG study need to be carefully and collectively considered, Including:
  1. (i)

    The conceptualization of the overall neuroscience problem

     
  2. (ii)

    Selection of the specific research question

     
  3. (iii)

    The rationale for choosing MEG to assess such brain function

     
  4. (iv)

    The experimental design

     
  5. (v)

    The acquisition setup

     
  6. (vi)

    Tailoring of the measurement protocol to your subject population

     
  7. (vii)

    The choice of methods to analyze the MEG data and for statistical testing

     
  8. (viii)

    The choice of journal and other means to disseminate your work.

     

While many of these components are common to noninvasive brain methods in general, the aspects (iv), (v), and (vii), in particular, require a lot of expertise. Below, we discuss some of the key considerations and pitfalls to keep in mind and have organized these into categories of those that are important before, during, and after measurements. Under each section, we relay concrete advice (in italics) to beginner MEG users from seasoned experts in the field, such as Riitta Hari (RH), Yury Shtyrov (YS), Riitta Salmelin (RS), Robert Oostenveld (RO), Veikko Jousmäki (VJ), as well as ourselves (MH and DL).

6.1 Stage 1: Before Any Measurements

6.1.1 Invest Time in Training and Education

A general factor that influences the quality of your study is of course your level of expertise and the degree to which you are familiar with MEG, electrophysiology in general, and neuroscience in particular. You should make sure you invest in training and education, so you have the basics of MEG in place before you get started.

Are you familiar with the basic signals and artifacts in MEG? If not, repeat some well-known studies of sensory systems and also collect spontaneous activity in 1–2 subjects, analyze the data, and examine the results with an experienced colleague. And stay home for a week and read a good MEG–EEG Primer; it will pay off later, (RH)

Spend time getting acquainted with your measured MEG data. Learn to look at and identify patterns in evoked responses and power spectral distributions. Learn how artifacts appear in the MEG sensor signals. Overlay experimental conditions and compare the patterns. You should have a good idea of what is going on in the brain (as measured by MEG) before you start the source analysis. Even if you end up using some very advanced analysis method whose result is difficult or impossible to verify by going back to original measured data, e.g., connectivity analysis, you should always include in the experimental design some ways, e.g., triggers indicating some distinct experimental conditions/states, that allow you to extract some basic evoked responses and/or power spectra and check that everything is in good shape and makes sense at least at that low level. (RS)

Learn from others and from examples, and practice your skills. Data analysis methods have moved on considerably (are moving on) and training material (books, course lectures) do not cover the most recent methods. Also the expertise of supervisors (who received their own training in the past) is not always up to date. To deal with this knowledge gap, you should interact with people that are doing analysis now or have done so recently, you should look at recent examples and not at 20 year old stuff, and practice your analysis skills on new datasets. Note that you should not overanalyze your own data with a plethora of fancy-looking methods: it is better to train yourself on shared datasets and to transfer the insights this gives to your own studies. (RO)

6.1.2 Determine if MEG and Your Research Question Is a Good Match

The principal critical factor for a successful MEG study is of course the research question itself and how well it can be studied at all with MEG. You should initially make sure your research question has a clear neuroscience component and specifically addresses human brain function in a way that makes MEG as an imaging methodology a meaningful tool for your work.

Think carefully what your research question is. If it mainly refers to accurate localization (and if you believe that is, in general, the main organizational principle of brain function), and you are not focusing on the earliest sensory responses, then MEG may not be the best tool for you. However, if your question is largely/primarily about timing and how different brain regions operate together to achieve functions, then go for it. (RS)

Are you comparing two conditions or two subject groups? Does some clear behavioural measure, e.g. reaction time, differentiate them? If not, MEG is unlikely to provide any extra information…at least not for the "relative novice MEG user". Also remember that if, e.g., a group of patients differs from the control group already in their early responses, not too much can be said about the normality of brain processes underlying later responses. (RH)

6.1.3 Designing Your Experiment

The next critical factor is the design of your study or experiment. Since the experimental design is what transforms the neuroscientific research question into a controlled form that allows you to test your research hypothesis by comparing within- and/or between-subject conditions, it is tightly linked to your research question. Hence, when you design your experiment, you should not only be clear about how your experimental design tests your neuroscientific question but also be explicit about how your experimental conditions and events are represented in terms of a brain activity and how these events may be extracted as discrete spatiotemporal spectral events in your MEG/EEG data.

Make a guess of the active brain areas. Are they at the reach of MEG (e.g. not too deep, not too close to each other, not too variable, not totally radial…). Think how they should be seen in EEG and make that recording as well. After all, MEG and EEG are the two sides of the same coin and the results should not be in contradiction. (RH)

Design your analysis strategy before designing the stimulation protocol. That means thinking of your stimuli/triggers and their timing beforehand. (YS)

Better have over- than under-specified stimulation protocol (e.g. trigger numbers): you can always collapse across a few stim types for analysis, but the opposite is more difficult • MEG is more flexible than fMRI in terms of what you can do, but better avoid fMRI-style multiple levels of contrasts subtracting blocks from one another. Consider clearly orthogonal/factorial design for GLM analysis, or, vice versa sufficient variability in stimulation to allow regression/correlation approaches.(YS)

Maximize the number of trials for the modality in question. Always include control conditions, the more the better – remember that the brain’s activation is dominated by physical, not cognitive stimulus features. Balance this against the overall recording time (my rule is 2 hrs max, ideally under 1h). (YS)

The very accurate time tracking of MEG allows one to design experiments with very varied, intermixed designs. The resulting data can be averaged in multiple ways that enables close-ups on different stimuli, tasks, different phases of tasks, and also artefact signals related to, e.g., eye or mouth movements. Furthermore, the same raw data yields evoked responses, modulation of cortical rhythms, connectivity estimates… Make use of that hugely versatile and complementary information. (RS)

Make sure to consider the existing MEG and EEG literature carefully when you design your experiment, so you have an understanding of exactly how your experiment contributes to what is already known. Understanding this already during the experimental design process will allow you to design a study that targets a “new unknown” more precisely, and will also allow you to identify who the audience interested in this topic is and where they publish, so you can plan where you best disseminate your results later. (DL)

6.2 Stage 2: During Measurements

6.2.1 Piloting: Technical Piloting and Piloting on Human Subjects

Once you have your experiment designed and the initial version of your setup in place, it is very useful to invest time to make pilot measurements, first in technical dry runs, then (always) on yourself and your colleagues, and finally on naïve participants or patients. Although you may experience this as an unnecessary and time-consuming detour and as something that could be overlooked, piloting is always a worthwhile and educational investment as it forces you to explicitly evaluate how well your research question can be examined with your choice of experimental design and research participant and also allows you to learn about the exact nature of the brain responses you will use for your analysis and statistical testing.

Learn to know your enemies in MEG recordings. MEG measurement has several caveats. It is important to learn and master the MEG recording procedure and know where it can go wrong. Both MEG device and subject may cause artefacts – some of them should be removed before the actual measurement, while others are removed after measurements, off-line. Once you know the details and how to double check for possible failures, you will acquire more reliable data. (VJ)

Make many pilot experiments and examine carefully their results before starting the real experiment. Pay keen attention to the replicability of the results. Instruct the subjects carefully and monitor their state/vigilance/performance throughout the experiment—a sleepy subject can spoil the data in a very unexpected manner. (RH)

Consider the areas you expect to be activated, and respective source localization difficulties. Then, consider adding EEG. In fact, always consider doing EEG unless there are specific reasons not to (kids, patients). (YS)

Consider the supine position – it’s great to keeping the head still. (YS)

Don’t hesitate to pilot a few extra participants to make sure that the task and the participants behave as intended. Make a habit of always piloting the task on yourself, then on someone else with expert insights in MEG and/or extensive experience from being a participant in MEG experiments, and finally also on some naïve participants. (DL)

Start with a very basic experiment. Simple, well documented experiments are the ones to start with. If you cannot reach the expected responses and results in simple experiments, you need to learn where it went wrong. (VJ)

6.2.2 The Research Participant

During piloting and also in subsequent measurements constituting the actual data to be analyzed and published, it is crucial to consider the research participants view, such as his/her understanding of, and experience from, the experiment. Even in the scenario that you have the perfect experiment lined up from the perspective of research question, tasks, and stimulation, it can easily fail completely if you do not consider and pilot the experiment also from the perspective of the research participant, including interviews of the research participant. Questions such as whether the participant understands the task instructions correctly, how the participant solves the experimental task (if any), to what degree the participant is preoccupied with irrelevant concerns, and whether the participant stays alert and awake (or not) during the experiment will all strongly influence and potentially contaminate your experimental manipulation. In the ideal scenario, the subject understands the tasks, performs it as planned, focuses on the task, and stays attentive and awake, while in the suboptimal scenario, some or all of these fail.

Make sure to always assess the participants understanding of the task, e.g. by asking him/her how they would instruct someone else to perform the task. Make it a habit to assess the participants attentiveness during the experiment, and don’t underestimate how boring and sleep-inducing your fantastic experiment might be. A poorly designed task can put a lot of people to sleep in less than 10 minutes, and a well-designed task can keep one focused for more than an hour. The best way to identify this, is to try the task yourself, on an experienced colleague, on a naïve participant, and to assess sleepiness and attentiveness throughout plus interview the participants after recordings. (DL)

Respect your subject. Subjects play a crucial role in your experiments. They can perform the task as requested although sometimes they totally miss the task. A magnetically shielded room may be a very unfamiliar place for them and they may move a lot. Instruct your subject carefully, pay attention to the behaviour and performance throughout the task, reinstruct if needed, and make notes about the measurement. (VJ)

Never assume that the participant will do the experimental task the way you intend, but instead make sure to evaluate the different ways there are to solve, cheat or violate the task you set up. If there’s a way to do the task in different ways or in ways that require less effort, you risk lousy and irrelevant data. (DL)

6.2.3 The Actual MEG Measurements

During your main measurements, it is essential to maintain a clear protocol for how you recruit, screen, instruct, and prepare participants, as well as a clear standard procedure for how you check, assess, and monitor the technical quality of the measurements as well as the participant’s sleepiness and behavior during all measurements.

It is also advisable to include some characteristic data from the head position measurements, if available. In addition, there are open software solutions, which can visualize the head position relative to the MEG helmet during the measurements. These tools are recommended to ensure that the subject’s head is located centered, all the way in the helmet, and without a head tilt or rotation. Consistent head positioning makes sensor-space data easier to interpret, e.g., for interhemispheric differences and approximate comparisons across subjects are also facilitated.

Noise measurements should be carried out regularly, at least every morning or preferably before each recording session starts. These noise data should be archived systematically in order to be able to trace down possible technical problems in the analysis phase. In addition data recorded from the shielded room void of a subject (empty room data) is later valuable for estimating the noise covariance matrix, needed in many source estimation techniques, especially those designed for analyzing ongoing activity. Any observations, including artifacts and noisy channels should be recorded in a lab notebook. Despite the availability of sophisticated electronic notebook solutions, a conventional notebook on paper is still a valuable aid because, e.g., manual drawings` of observations can be easily included. (MH)

In many cases, averages to at least some of the events can be computed on line. This is an additional tool for data quality assurance and should be employed whenever possible. In addition, there are current real-time software projects, which aim at providing more complex analysis results, even source estimates, during the recording. When available, these tools are highly recommended as well provided that they do not distract the experimenter from observing the subject and the incoming raw data. (MH)

The rules at many imaging centers require that two persons are present during each measurement. While this is a safeguard against consequences of some unexpected adverse event, under normal conditions, the online quality assurance tasks can be divided between the two persons present.

6.3 Stage 3: After Completed Measurements

Often, a fully detailed analysis of data is done only when all data is already collected, although it is very useful to make sure you develop a strategy for your analysis as well as scripts at least for the fundamental analyzes you plan already when you design and pilot your experiment.

6.3.1 Analysis

Analysis methods are neither omnipotent nor error-free (they are made by humans, after all), so check the results every step of the way (mastering step 3 is an immense help here!). A method that seems to have worked very well on some data set may have surprising (unwanted) effects on another data set, where the experimental design and non-brain disturbances are different. The more contrived the method is and the less intervention it requires from the user (because it inherently assumes so much), the more alert you should be. (RS)

Understand the effects that confounds have on the signal, signal processing results and statistics. It is tempting (and career-wise tempting) to report the desired interpretation, but (scientifically) important to report the correct interpretation of your findings. With experimental designs and analyses getting more complex (i.e. moving on beyond what was done 20 years ago), it is not always obvious which non-interesting interpretations (confounds) can provide alternative interpretations of the results. Testing alternative hypotheses and interpretations of results is key to doing good science. (RO)

Do not try to do the most sophisticated analysis if you haven’t mastered the basic analyses yet. Although it is tempting to do the most fancy analyses, these are not always well established. If you want to use sophisticated methods, you should realize that this requires a serious effort in learning the possibilities, limitations and pitfalls of these methods. Without a good understanding of the methods, chances are that you will be contributing to the ever growing pool of bogus (false positive, non-reproducible, and incorrectly interpreted) results that happen to be published. (RO)

Learn your enemies in source modelling. Browse through the data for unexpected artefacts before post-processing steps. Try to visualize the data before using any black-box approach – if garbage goes in garbage will come out. Once again, start to analyse a simple data set to verify your analysis setup before entering the world of more complicated responses possible with a lower signal-to-noise level. (VJ)

Aim to use multiple analysis methods with different underlying assumptions to address the same question, such as ECD and MNE to find cortical mapping of evoked responses. This can importantly help you to avoid misinterpretations, e.g., due to artefact signals. After all, we all want the results to be as true as possible, and not merely a reflection of a certain analysis method. (RS)

Enjoy your recordings. Once you know the caveats it is easier to move on to the more demanding tasks and analysis methods. Familiarize yourself with a given sensory modality and existing evidence in literature. Experiment with your task, stimuli, and data acquisition settings carefully before commencing the actual experiments in MEG in subjects or patients. (VJ)

Since rigor and reproducibility benefits significantly from sharing your analysis scripts and data (Jas et al. 2018), it is highly recommended that you plan and annotate your analysis scripts as well as pseudonymize and organize your data so that you can later share these with other researchers. Of course, for this to be possible and legal in practice, you need to consider such dissemination early on so that it is covered in your ethical permit and in the participant’s written consent.

6.3.2 Sharing Data

Share your data and analyses. Knowledge and understanding is only built up in an open scientific discourse. With complex and data-heavy empirical studies, sharing theories and interpretations of results is not sufficient any more, since too many relevant details get lost in translation. Full disclosure of data, analysis methods (e.g. scripts and software) and results are required to complement the traditional publications. The published data and methods contributes to increased reproducibility and replicability, and helps others to learn from and build upon your research. (RO)

7 Summary

In the previous sections, we have outlined the historical role of MEG in the context of neuroimaging methods (Sect. 1), outlined what aspects of brain activity that MEG and EEG may measure (Sect. 2), presented different approaches to how such measures may be used to draw conclusions about the underlying neuronal source (Sect. 3), discussed the past and future of MEG and EEG instrumentation (Sect. 4), deliberated on the usefulness of MEG and EEG both in terms of academic influence (Sect. 5), and finally presented basic advice to the novice user (Sect. 6).

In year 2018, MEG celebrated the 50th anniversary of the initial measurements of alpha activity by David Cohen. Since then, there has been a steady and increasing contribution to the advancement of our understanding of the brains functional organization, by a continuing development of instrumentation, analysis methods, applications, and academic influence.

At its heart, science is an iterative process of knowledge improvement, where no detail of knowledge is ever enough. This is also true for MEG, not only in theory but also in practice: exciting development in sensor technology, methods, and applications gives encouraging promise to the usefulness of MEG in advancing the understanding of the brain’s functional organization for decades to come.

References

  1. Adrian D (1944) Brain rhythms. Nature 153:360–362CrossRefGoogle Scholar
  2. Agam Y et al.(2011) Multimodal neuroimaging dissociates hemodynamic and electrophysiological correlates of error processing. Proc Natl Acad Sci U S A 108:17556–17561. https://doi.org/10.1073/pnas.1103475108CrossRefGoogle Scholar
  3. Ahlfors S, Ilmoniemi R, Hämäläinen M (1992) Estimates of visually evoked cortical currents. Electroencephalogr Clin Neurophysiol 82:225–236CrossRefGoogle Scholar
  4. Ahlfors SP, Han J, Lin FH, Witzel T, Belliveau JW, Hamalainen MS, Halgren E (2010) Cancellation of EEG and MEG signals generated by extended and distributed sources. Hum Brain Mapp 31:140–149. https://doi.org/10.1002/hbm.20851CrossRefGoogle Scholar
  5. Ahveninen J et al.(2006) Task-modulated "what" and "where" pathways in human auditory cortex. Proc Natl Acad Sci U S A 103:14608–14613CrossRefGoogle Scholar
  6. Aine C et al.(1996) Retinotopic organization of human visual cortex: Departures from the classical model. Cereb Cortex 6:354–361CrossRefGoogle Scholar
  7. Alexander GE, DeLong MR, Strick PL (1986) Parallel organization of functionally segregated circuits linking basal ganglia and cortex. Annu Rev Neurosci 9:357–381. https://doi.org/10.1146/annurev.ne.09.030186.002041CrossRefGoogle Scholar
  8. Andersen LM et al.(2017) Similarities and differences between on-scalp and conventional in-helmet magnetoencephalography recordings. PLoS One 12:e0178602. https://doi.org/10.1371/journal.pone.0178602CrossRefGoogle Scholar
  9. Babadi B, Obregon-Henao G, Lamus C, Hamalainen MS, Brown EN, Purdon PL (2014) A Subspace Pursuit-based Iterative Greedy Hierarchical solution to the neuromagnetic inverse problem. NeuroImage 87:427–443. https://doi.org/10.1016/j.neuroimage.2013.09.008CrossRefGoogle Scholar
  10. Baillet S (2017) Magnetoencephalography for brain electrophysiology and imaging. Nat Neurosci 20:327–339. https://doi.org/10.1038/nn.4504CrossRefGoogle Scholar
  11. Baillet S, Mosher JC, Leahy RM (2001) Electromagnetic Brain Mapping. IEEE Signal Process Mag 18:14–30CrossRefGoogle Scholar
  12. Barry JF et al.(2016) Optical magnetic detection of single-neuron action potentials using quantum defects in diamond. Proc Natl Acad Sci U S A 113:14133–14138. https://doi.org/10.1073/pnas.1601513113CrossRefGoogle Scholar
  13. Berger H (1929) Ueber das Elektrenkephalogramm des Menschen Archiv fuer Psychiatrie und Nervenkrankheiten, vol 87, pp 527–570Google Scholar
  14. Borna A et al.(2017) A 20-channel magnetoencephalography system based on optically pumped magnetometers. Phys Med Biol 62:8909–8923. https://doi.org/10.1088/1361-6560/aa93d1CrossRefGoogle Scholar
  15. Boto E et al.(2017) A new generation of magnetoencephalography: Room temperature measurements using optically-pumped magnetometers. NeuroImage 149:404–414. https://doi.org/10.1016/j.neuroimage.2017.01.034CrossRefGoogle Scholar
  16. Boto E et al.(2018) Moving magnetoencephalography towards real-world applications with a wearable system. Nature 555:657–661. https://doi.org/10.1038/nature26147CrossRefGoogle Scholar
  17. Brenner D, Williamson SJ, Kaufman L (1975) Visually evoked magnetic fields of the human brain. Science 190:480–481CrossRefGoogle Scholar
  18. Canolty RT et al.(2006) High gamma power is phase-locked to theta oscillations in human neocortex. Science 313:1626–1628. https://doi.org/10.1126/science.1128115CrossRefGoogle Scholar
  19. Cichy RM, Pantazis D, Oliva A (2014) Resolving human object recognition in space and time. Nat Neurosci 17:455–462. https://doi.org/10.1038/nn.3635CrossRefGoogle Scholar
  20. Cichy RM, Pantazis D, Oliva A (2016) Similarity-Based Fusion of MEG and fMRI Reveals Spatio-Temporal Dynamics in Human Cortex During Visual Object Recognition. Cereb Cortex 26:3563–3579. https://doi.org/10.1093/cercor/bhw135CrossRefGoogle Scholar
  21. Cohen D (1968) Magnetoencephalography: evidence of magnetic fields produced by alpha-rhythm currents. Science 161:784–786CrossRefGoogle Scholar
  22. Cohen D (1972) Magnetoencephalography: detection of the brain's electrical activity with a superconducting magnetometer. Science 175:664–666CrossRefGoogle Scholar
  23. Cuffin BN, Cohen D (1977) Magnetic fields of a dipole in special volume conductor shapes. IEEE Trans Biomed Eng 24:372–381CrossRefGoogle Scholar
  24. Dalal SS et al.(2009) Simultaneous MEG and intracranial EEG recordings during attentive reading. NeuroImage 45:1289–1304. https://doi.org/10.1016/j.neuroimage.2009.01.017CrossRefGoogle Scholar
  25. Dale AM, Sereno MI (1993) Improved localization of cortical activity by combining EEG and MEG with MRI cortical surface reconstruction: A linear approach. J Cogn Neurosci 5:162–176CrossRefGoogle Scholar
  26. Dale AM, Liu AK, Fischl BR, Buckner RL, Belliveau JW, Lewine JD, Halgren E (2000) Dynamic statistical parametric mapping: combining fMRI and MEG for high-resolution imaging of cortical activity. Neuron 26:55–67CrossRefGoogle Scholar
  27. Dannhauer M, Lanfer B, Wolters CH, Knosche TR (2011) Modeling of the human skull in EEG source analysis. Hum Brain Mapp 32:1383–1399. https://doi.org/10.1002/hbm.21114CrossRefGoogle Scholar
  28. Daunizeau J et al.(2007a) Assessing the relevance of fMRI-based prior in the EEG inverse problem: a bayesian model comparison approach. IEEE Trans Signal Process 53:3461–3472MathSciNetCrossRefGoogle Scholar
  29. Daunizeau J et al.(2007b) Symmetrical event-related EEG/fMRI information fusion in a variational Bayesian framework. NeuroImage 36:69–87. https://doi.org/10.1016/j.neuroimage.2007.01.044CrossRefGoogle Scholar
  30. Drechsler F, Wolters CH, Dierkes T, Si H, Grasedyck L (2009) A full subtraction approach for finite element method based source analysis using constrained Delaunay tetrahedralisation. NeuroImage 46:1055–1065. https://doi.org/10.1016/j.neuroimage.2009.02.024CrossRefGoogle Scholar
  31. Faley MI, Poppe U, Borkowski RE, Schiek M, Boers F (2012) Magnetoencephalography using a Multilayer hightc DC SQUID Magnetometer. Phys Procedia 36:66–71CrossRefGoogle Scholar
  32. Goldenholz DM, Ahlfors SP, Hamalainen MS, Sharon D, Ishitobi M, Vaina LM, Stufflebeam SM (2009) Mapping the signal-to-noise-ratios of cortical sources in magnetoencephalography and electroencephalography. Hum Brain Mapp 30:1077–1086. https://doi.org/10.1002/hbm.20571CrossRefGoogle Scholar
  33. Gramfort A, Kowalski M, Hamalainen M (2012) Mixed-norm estimates for the M/EEG inverse problem using accelerated gradient methods. Phys Med Biol 57:1937–1961. https://doi.org/10.1088/0031-9155/57/7/1937CrossRefGoogle Scholar
  34. Gramfort A, Strohmeier D, Haueisen J, Hamalainen MS, Kowalski M (2013) Time-frequency mixed-norm estimates: sparse M/EEG imaging with non-stationary source activations. NeuroImage 70:410–422. https://doi.org/10.1016/j.neuroimage.2012.12.051CrossRefGoogle Scholar
  35. Graybiel AM (2000) The basal ganglia. Curr Biol CB 10:R509–R511CrossRefGoogle Scholar
  36. Grinvald A, Hildesheim R (2004) VSDI: a new era in functional imaging of cortical dynamics. Nat Rev Neurosci 5:874–885. https://doi.org/10.1038/nrn1536CrossRefGoogle Scholar
  37. Gullmar D, Haueisen J, Reichenbach JR (2010) Influence of anisotropic electrical conductivity in white matter tissue on the EEG/MEG forward and inverse solution. A high-resolution whole head simulation study. NeuroImage 51:145–163. https://doi.org/10.1016/j.neuroimage.2010.02.014CrossRefGoogle Scholar
  38. Hämäläinen M, Hari R (2002) Magnetoencephalographic characterization of dynamic brain activation: basic principles and methods of data collection and source analysis. In: Toga AW, Mazziotta JC (eds) Brain Mapping, The Methods. Academic Press, San Diego, pp 227–253CrossRefGoogle Scholar
  39. Hämäläinen M, Ilmoniemi R (1984) Interpreting magnetic fields of the brain: minimum norm estimates. Helsinki University of Technology, EspooGoogle Scholar
  40. Hämäläinen MS, Sarvas J (1987) Feasibility of the homogeneous head model in the interpretation of neuromagnetic fields. Phys Med Biol 32:91–97CrossRefGoogle Scholar
  41. Hämäläinen MS, Sarvas J (1989) Realistic conductivity geometry model of the human head for interpretation of neuromagnetic data. IEEE Trans Biomed Eng BME 36:165–171CrossRefGoogle Scholar
  42. Hämäläinen M, Hari R, Ilmoniemi R, Knuutila J, Lounasmaa OV (1993) Magnetoencephalography – theory, instrumentation, and applications to noninvasive studies of the working human brain. Rev Mod Phys 65:413–497CrossRefGoogle Scholar
  43. Hari R, Ilmoniemi RJ (1986) Cerebral magnetic fields. Crit Rev Biomed Eng 14:93–126Google Scholar
  44. Hari R, Kaukoranta E (1985) Neuromagnetic studies of the somatosensory system. Progr Neurobiol 24:233–256CrossRefGoogle Scholar
  45. Hari R, Salmelin R (2012) Magnetoencephalography: from SQUIDs to neuroscience. Neuroimage 20th anniversary special edition. NeuroImage 61:386–396. https://doi.org/10.1016/j.neuroimage.2011.11.074CrossRefGoogle Scholar
  46. Heilbronner SR, Platt ML (2013) Causal evidence of performance monitoring by neurons in posterior cingulate cortex during learning. Neuron 80:1384–1391. https://doi.org/10.1016/j.neuron.2013.09.028CrossRefGoogle Scholar
  47. Holmes N et al.(2018) A bi-planar coil system for nulling background magnetic fields in scalp mounted magnetoencephalography. NeuroImage 181:760–774. https://doi.org/10.1016/j.neuroimage.2018.07.028CrossRefGoogle Scholar
  48. Horwitz B, Poeppel D (2002) How can EEG/MEG and fMRI/PET data be combined? Hum Brain Mapp 17:1–3. https://doi.org/10.1002/hbm.10057CrossRefGoogle Scholar
  49. Iivanainen J, Stenroos M, Parkkonen L (2017) Measuring MEG closer to the brain: Performance of on-scalp sensor arrays. NeuroImage 147:542–553. https://doi.org/10.1016/j.neuroimage.2016.12.048CrossRefGoogle Scholar
  50. Ilmoniemi RJ, Hämäläinen MS, Knuutila J (1985) The forward and inverse problems in the spherical model. In: Weinberg H, Stroink G, Katila T (eds) Biomagnetism: Applications & Theory. Pergamon Press, New York, pp 278–282Google Scholar
  51. Jas M, Larson E, Engemann DA, Leppakangas J, Taulu S, Hamalainen M, Gramfort A (2018) A Reproducible MEG/EEG Group Study With the MNE Software: Recommendations, Quality Assessments, and Good Practices. Front Neurosci 12:530. https://doi.org/10.3389/fnins.2018.00530CrossRefGoogle Scholar
  52. Jones EG (1998) Viewpoint: the core and matrix of thalamic organization. Neuroscience 85:331–345CrossRefGoogle Scholar
  53. Jones EG (2001) The thalamic matrix and thalamocortical synchrony. Trends Neurosci 24:595–601CrossRefGoogle Scholar
  54. Jones SR, Pritchett DL, Stufflebeam SM, Hamalainen M, Moore CI (2007) Neural correlates of tactile detection: a combined magnetoencephalography and biophysically based computational modeling study. J Neurosci Off J Soc Neurosci 27:10751–10764CrossRefGoogle Scholar
  55. Jones SR, Pritchett DL, Sikora MA, Stufflebeam SM, Hamalainen M, Moore CI (2009) Quantitative analysis and biophysically realistic neural modeling of the MEG mu rhythm: rhythmogenesis and modulation of sensory-evoked responses. J Neurophysiol 102:3554–3572. https://doi.org/10.1152/jn.00535.2009CrossRefGoogle Scholar
  56. Kakisaka Y et al.(2012) Use of simultaneous depth and MEG recording may provide complementary information regarding the epileptogenic region. Epileptic Disord 14:298–303. https://doi.org/10.1684/epd.2012.0517CrossRefGoogle Scholar
  57. Kemppainen P, Ilmoniemi RJ (1989) Channel capacity of multichannel magnetometers. In: Williamson SJ, Hoke M, Stroink G, Kotani M (eds) Advances in Biomagnetism. Plenum Press, New York, pp 635–639CrossRefGoogle Scholar
  58. Kerr CE et al.(2011) Effects of mindfulness meditation training on anticipatory alpha modulation in primary somatosensory cortex. Brain Res Bull 85:96–103. https://doi.org/10.1016/j.brainresbull.2011.03.026CrossRefGoogle Scholar
  59. Khan S et al.(2013) Local and long-range functional connectivity is reduced in concert in autism spectrum disorders. Proc Natl Acad Sci U S A 110:3107–3112. https://doi.org/10.1073/pnas.1214533110CrossRefGoogle Scholar
  60. Kim K, Begus S, Xia H, Lee SK, Jazbinsek V, Trontelj Z, Romalis MV (2014) Multi-channel atomic magnetometer for magnetoencephalography: a configuration study. NeuroImage 89:143–151. https://doi.org/10.1016/j.neuroimage.2013.10.040CrossRefGoogle Scholar
  61. Knuutila J et al.(1991) Design of a 122-channel neuromagnetometer covering the whole head. In: Hoke M (ed) Abstract book of the 8th international conference on biomagnetism. Westfälische Wilhelms-Universität, Münster, pp 109–110Google Scholar
  62. Kominis IK, Kornack TW, Allred JC, Romalis MV (2003) A subfemtotesla multichannel atomic magnetometer. Nature 422:596–599. https://doi.org/10.1038/nature01484CrossRefGoogle Scholar
  63. Kriegeskorte N, Kievit RA (2013) Representational geometry: integrating cognition, computation, and the brain. Trends Cogn Sci 17:401–412. https://doi.org/10.1016/j.tics.2013.06.007CrossRefGoogle Scholar
  64. Kriegeskorte N, Mur M, Bandettini P (2008) Representational similarity analysis - connecting the branches of systems neuroscience. Front Syst Neurosci 2:4. https://doi.org/10.3389/neuro.06.004.2008CrossRefGoogle Scholar
  65. Krishnaswamy P et al.(2017) Sparsity enables estimation of both subcortical and cortical activity from MEG and EEG. Proc Natl Acad Sci U S A 114:E10465–E10474CrossRefGoogle Scholar
  66. Lanfer B, Wolters CH, Demokritov SO, Pantev C (2007) Validating finite element method based EEG and MEG forward computations. Paper presented at the 41 Jahrestagung der Deutschen Gesellschaft für Biomedizinische Technik AachenGoogle Scholar
  67. Lew S, Wolters CH, Dierkes T, Röer C, MacLeod RS (2009) Accuracy and run-time comparison for different potential approaches and iterative solvers in finite element method based EEG source analysis. Appl Numer Math 59:1970–1988MathSciNetCrossRefGoogle Scholar
  68. Lew S, Sliva DD, Choe MS, Grant PE, Okada Y, Wolters CH, Hamalainen MS (2013) Effects of sutures and fontanels on MEG and EEG source analysis in a realistic infant head model. NeuroImage 76C:282–293. https://doi.org/10.1016/j.neuroimage.2013.03.017CrossRefGoogle Scholar
  69. Liu AK, Belliveau JW, Dale AM (1998) Spatiotemporal imaging of human brain activity using functional MRI constrained magnetoencephalography data: Monte Carlo simulations. Proc Natl Acad Sci U S A 95:8945–8950CrossRefGoogle Scholar
  70. Lopes da Silva F, Van Rotterdam A (1992) Biophysical aspects of EEG and MEG generation. In: Niedermeyer E, Lopes da Silva F (eds) Electroencephalography: basic principles, clinical applications and related fields. W&W, Baltimore/MunichGoogle Scholar
  71. Mosher JC, Leahy RM (1998) Recursive MUSIC: a framework for EEG and MEG source localization. IEEE Trans Biomed Eng 45:1342–1354CrossRefGoogle Scholar
  72. Mosher JC, Lewis PS, Leahy RM (1992) Multiple dipole modeling and localization from spatio-temporal MEG data. IEEE Trans Biomed Eng 39:541–557. https://doi.org/10.1109/10.141192CrossRefGoogle Scholar
  73. Mosher JC, Leahy RM, Lewis PS (1999) EEG and MEG: forward solutions for inverse methods. IEEE Trans Biomed Eng 46:245–259CrossRefGoogle Scholar
  74. Murakami S, Okada Y (2015) Invariance in current dipole moment density across brain structures and species: physiological constraint for neuroimaging. NeuroImage 111:49–58. https://doi.org/10.1016/j.neuroimage.2015.02.003CrossRefGoogle Scholar
  75. Murakami H et al.(2016) Correlating magnetoencephalography to stereo-electroencephalography in patients undergoing epilepsy surgery. Brain J Neurol. https://doi.org/10.1093/brain/aww215CrossRefGoogle Scholar
  76. Nishitani N, Avikainen S, Hari R (2004) Abnormal imitation-related cortical activation sequences in Asperger's syndrome. Ann Neurol 55:558–562. https://doi.org/10.1002/ana.20031CrossRefGoogle Scholar
  77. Okada Y, Lahteenmaki A, Xu C (1999) Comparison of MEG and EEG on the basis of somatic evoked responses elicited by stimulation of the snout in the juvenile swine Clinical neurophysiology: official journal of the International Federation of Clinical. Neurophysiology 110:214–229CrossRefGoogle Scholar
  78. Okada Y et al.(2016) BabyMEG: A whole-head pediatric magnetoencephalography system for human brain development research. Rev Sci Instrum 87:094301. https://doi.org/10.1063/1.4962020CrossRefGoogle Scholar
  79. Ou W, Hamalainen MS, Golland P (2009) A distributed spatio-temporal EEG/MEG inverse solver. NeuroImage 44:932–946. https://doi.org/10.1016/j.neuroimage.2008.05.063CrossRefGoogle Scholar
  80. Ou W, Nummenmaa A, Ahveninen J, Belliveau JW, Hamalainen MS, Golland P (2010) Multimodal functional imaging using fMRI-informed regional EEG/MEG source estimation. NeuroImage 52:97–108. https://doi.org/10.1016/j.neuroimage.2010.03.001CrossRefGoogle Scholar
  81. Phelps EA, LeDoux JE (2005) Contributions of the amygdala to emotion processing: from animal models to human behavior. Neuron 48:175–187. https://doi.org/10.1016/j.neuron.2005.09.025CrossRefGoogle Scholar
  82. Plonsey R (1969) Bioelectric phenomena. McGraw-Hill, New YorkGoogle Scholar
  83. Roche-Labarbe N, Aarabi A, Kongolo G, Gondry-Jouet C, Dumpelmann M, Grebe R, Wallois F (2008) High-resolution electroencephalography and source localization in neonates. Hum Brain Mapp 29:167–176. https://doi.org/10.1002/hbm.20376CrossRefGoogle Scholar
  84. Sacchet MD et al.(2015) Attention drives synchronization of alpha and beta rhythms between right inferior frontal and primary sensory neocortex. J Neurosci Off J Soc Neurosci 35:2074–2082. https://doi.org/10.1523/JNEUROSCI.1292-14.2015CrossRefGoogle Scholar
  85. Salmelin R, Hari R (1994) Spatiotemporal characteristics of sensorimotor neuromagnetic rhythms related to thumb movement. Neuroscience 60:537–550CrossRefGoogle Scholar
  86. Salmelin R, Hari R, Lounasmaa OV, Sams M (1994) Dynamics of brain activation during picture naming. Nature 368:463–465CrossRefGoogle Scholar
  87. Sams M, Hari R (1991) Magnetoencephalography in the study of human auditory information processing. Ann N Y Acad Sci 620:102–117CrossRefGoogle Scholar
  88. Sander TH, Preusser J, Mhaskar R, Kitching J, Trahms L, Knappe S (2012) Magnetoencephalography with a chip-scale atomic magnetometer. Biomed Opt Express 3:981–990. https://doi.org/10.1364/BOE.3.000981CrossRefGoogle Scholar
  89. Sarvas J (1987) Basic mathematical and electromagnetic concepts of the biomagnetic inverse problem. Phys Med Biol 32:11–22CrossRefGoogle Scholar
  90. Sekihara K, Nagarajan SS (2008) Adaptive spatial filters for electromagnetic brain imaging. Springer, BerlinGoogle Scholar
  91. Sharon D, Hamalainen MS, Tootell RB, Halgren E, Belliveau JW (2007) The advantage of combining MEG and EEG: Comparison to fMRI in focally stimulated visual cortex. NeuroImage 36:1225–1235CrossRefGoogle Scholar
  92. Sherman MA et al.(2016) Neural mechanisms of transient neocortical beta rhythms: Converging evidence from humans, computational modeling, monkeys, and mice Proceedings of the. Natl Acad Sci USA 113:E4885–E4894. https://doi.org/10.1073/pnas.1604135113CrossRefGoogle Scholar
  93. Steinstrater O, Sillekens S, Junghoefer M, Burger M, Wolters CH (2010) Sensitivity of beamformer source analysis to deficiencies in forward modeling. Hum Brain Mapp 31:1907–1927. https://doi.org/10.1002/hbm.20986CrossRefGoogle Scholar
  94. Stenroos M (2016) Integral equations and boundary-element solution for static potential in a general piece-wise homogeneous volume conductor. Phys Med Biol 61:N606–N617. https://doi.org/10.1088/0031-9155/61/22/N606CrossRefGoogle Scholar
  95. Stenroos M, Hauk O (2013) Minimum-norm cortical source estimation in layered head models is robust against skull conductivity error. NeuroImage 81:265–272. https://doi.org/10.1016/j.neuroimage.2013.04.086CrossRefGoogle Scholar
  96. Stenroos M, Nummenmaa A (2016) Incorporating and Compensating Cerebrospinal Fluid in Surface-Based Forward Models of Magneto- and Electroencephalography. PLoS One 11:e0159595. https://doi.org/10.1371/journal.pone.0159595CrossRefGoogle Scholar
  97. Stenroos M, Hunold A, Haueisen J (2014) Comparison of three-shell and simplified volume conductor models in magnetoencephalography. NeuroImage 94:337–348. https://doi.org/10.1016/j.neuroimage.2014.01.006CrossRefGoogle Scholar
  98. Steriade M, McCormick DA, Sejnowski TJ (1993) Thalamocortical oscillations in the sleeping and aroused brain. Science 262:679–685CrossRefGoogle Scholar
  99. Tanaka N et al.(2010) Propagation of epileptic spikes reconstructed from spatiotemporal magnetoencephalographic and electroencephalographic source analysis. NeuroImage 50:217–222. https://doi.org/10.1016/j.neuroimage.2009.12.033CrossRefGoogle Scholar
  100. Taylor JM et al.(2008) High-sensitivity diamond magnetometer with nanoscale resolution. Nat Phys 4:810–816. https://doi.org/10.1038/nphys1075CrossRefGoogle Scholar
  101. Tripp JH (1983) Physical concepts and mathematical models. In: Williamson SJ, Romani GL, Kaufman L, Modena I (eds) Biomagnetism: An Interdisciplinary Approach. Plenum, New York, pp 101–139CrossRefGoogle Scholar
  102. Tuch DS, Wedeen VJ, Dale AM, George JS, Belliveau JW (2001) Conductivity tensor mapping of the human brain using diffusion tensor. MRI Proc Natl Acad Sci USA 98:11697–11701CrossRefGoogle Scholar
  103. Tuomisto T, Hari R, Katila T, Poutanen T, Varpula T (1983) Studies of auditory evoked magnetic and electric responses: modality specificity and modelling. Il Nuovo Cimento 2D:471–494CrossRefGoogle Scholar
  104. Uutela K, Hämäläinen M, Somersalo E (1999) Visualization of magnetoencephalographic data using minimum current estimates. NeuroImage 10:173–180CrossRefGoogle Scholar
  105. Van Veen B, Buckley K (1988) Beamforming: A versatile approach to spatial filtering IEEE assp magazineGoogle Scholar
  106. Vesanen PT et al.(2013) Hybrid ultra-low-field MRI and magnetoencephalography system based on a commercial whole-head neuromagnetometer. Magn Reson Med 69:1795–1804. https://doi.org/10.1002/mrm.24413CrossRefGoogle Scholar
  107. Wan Q, Kerr C, Pritchett D, Hamalainen M, Moore C, Jones S (2011) Dynamics of dynamics within a single data acquisition session: variation in neocortical alpha oscillations in human MEG. PLoS One 6:e24941. https://doi.org/10.1371/journal.pone.0024941CrossRefGoogle Scholar
  108. Wang C, Sun L, Lichtenwalter B, Zerkle B, Okada Y (2016) Compact, ultra-low vibration, closed-cycle helium recycler for uninterrupted operation of MEG with SQUID magnetometers. Cryogenics 76:16–22CrossRefGoogle Scholar
  109. Williamson SJ, Kaufman L (1981) Biomagnetism. J Magn Magn Mat 22:129–202CrossRefGoogle Scholar
  110. Wolters CH, Anwander A, Berti G, Hartmann U (2007a) Geometry-adapted hexahedral meshes improve accuracy of finite-element-method-based EEG source analysis. IEEE Trans Biomed Eng 54:1446–1453. https://doi.org/10.1109/TBME.2007.890736CrossRefGoogle Scholar
  111. Wolters CH, Köstler H, Möller C, Härdtlein J, Anwander A (2007b) Numerical approaches for dipole modeling in finite element method based source analysis. Int Congr Ser 1300:189–192. ElsevierCrossRefGoogle Scholar
  112. Wolters CH, Köstler H, Möller C, Härtlein J, Grasedyck L, Hackbusch W (2007c) Numerical mathematics of the subtraction method for the modeling of a current dipole in EEG source reconstruction using finite element head models. SIAM J Sci Comput 30:24–45MathSciNetCrossRefGoogle Scholar
  113. Xie M et al.(2017) Benchmarking for On-Scalp MEG Sensors. IEEE Trans Biomed Eng 64:1270–1276. https://doi.org/10.1109/TBME.2016.2599177CrossRefGoogle Scholar

Copyright information

© Springer Nature Switzerland AG 2019

Authors and Affiliations

  1. 1.Athinoula A. Martinos Center for Biomedical ImagingMassachusetts General HospitalCharlestownUSA
  2. 2.Harvard Medical SchoolBostonUSA
  3. 3.NatMEG, Department of Clinical NeuroscienceKarolinska InstitutetStockholmSweden

Section editors and affiliations

  • Seppo P. Ahlfors
    • 1
    • 2
  1. 1.Department of Radiology, MGH/HST Athinoula A. Martinos Center for Biomedical ImagingMassachusetts General HospitalCharlestown, MAUSA
  2. 2.Harvard-MIT Division of Health Sciences and TechnologyCambridgeUSA

Personalised recommendations