|Home | About | Journals | Submit | Contact Us | Français|
For presurgical evaluation of resection of a mass or other lesion, noninvasive neuroimaging aids in the surgical planning and in counseling patients about possible risks of the surgery. MEG performs the most common types of surgical planning that the neurosurgeon faces, including localization of epileptic discharges, determination of the hemispheric dominance of verbal processing, and the ability to locate eloquent cortex. MEG is most useful when it is combined with structural imaging, most commonly combined with structural MRI and MR diffusion imaging. We review the history of clinical MEG, introduce the basic concepts about biophysics of MEG, and outline the basic neurosurgical applications of MEG.
In 1968, David Cohen (see Figure 1) used a room-temperature copper coil as a detector to recorded the first magnetoencephalogram at the university of Illinois . Later, at the Massachusetts Institute of Technology, he built a more elaborate shielded room. At about the same time, James Zimmerman and colleagues developed the Superconducting Quantum Interference Device (SQUID), which uses the Josephson junction to measure tiny magnetic fields. It requires cooling to liquid helium temperatures, and has a sensitivity of several hundreds of times that of a copper coil. Zimmerman brought this detector to Cohen's room, and this combination of shielding and detector allowed the first clear measurements of the body's magnetic fields. After they measured the heart, Cohen next recorded the first MEG measured with a SQUID .
For the initial measurements of magnetic brain activity, physicists and neuroscientists used a single or a just few magnetometers. Localizing activity with such a low number of sensors required moving them in order to sampling from various locations over the head. This was time consuming and not practical for routine clinical use. Clinical MEG became possible when MEG systems were developed that provided coverage of approximately 7–12 cm, enough to produce a field map large enough to visualize the magnetic field lines from a single dipole source. This made it possible to localize the activity in a registered brain MRI.
Modern MEG systems now provide hundreds of channels that can provide whole-head coverage (See Figure 2). This makes it possible to map activity throughout the cerebral cortex and is critical for presurgical mapping of language areas, and for detecting propagating or widespread epileptic activity. Due to interference from extraneous magnetic fields, all MEG measurements must be performed in a magnetically shielded room, which typically consists of two to four layers of aluminum and multiple layers of ferromagnetic shielding.
Unlike other hemodynamic techniques (fMRI and PET), both MEG and electroencephalography (EEG) directly measures electric brain activity. The neural generators of the MEG and EEG are identical. There are critical differences that make them both useful and complementary. MEG preferentially detects activity in superficial, non-radial areas of cortex, such as the fissural cortex of the cerebral hemispheres. This is particularly advantageous if the area of the area of interest is also radial, such as the primary somatosensory cortex, which lies in the walls of the sulci.
Much of the neural activity measured by MEG originates as the post-synaptic activity in the pyramidal cells of the cerebral cortex. MEG measures the vector sum of post-synaptic potentials, as contrasted with BOLD fMRI and some form of PET imaging that reflects neural activity indirectly through changes in blood flow. MEG localizes neural activity more accurately than EEG because magnetic fields are less perturbed than electrical potentials by overlying brain structures: scalp, skull, cerebrospinal fluid, meninges, and vascular structures. Recently, statistical combination of structural MRI, functional MRI, with MEG have taken a great stride forward by yielding the maximum benefit from each technique into a single image [3, 4].
The calculation of the magnetic field is more straightforward than that of the electric field because of the symmetries and conductivity distribution of the human head. Since the EEG also is influenced by the extracellular volume current, which are difficult to model accurately. All currents, both intracellular and extracellular, generate magnetic fields, but, due to the near spherical shape of the head, one can calculate the resultant magnetic fields due to primary currents without taking into account the conductivity layers of the head.
Intelligent interpretation of MEG requires the examination of the waveforms recorded by the SQUIDS sensors, followed by a source estimate that reflects the site of activity in the brain. The raw measurement of MEG is a time varying magnetic field. For epilepsy discharge mapping, segments of the activity is used for source analysis, often without any signal averaging. In order to improve the signal-to-noise ratio of measured signals for evoked activity, such as language laterality, however, it is often necessary to average several (typically around 100 trials) responses from identical or similar stimuli. The signal is typically averaged based on the timing of the presented stimulus, the subject's response, or the peak of activity, such as the peak of an epileptic spike. Averaging lowers the effects of extraneous activity, effectively improving the overall signal-to-noise ratio of the desired activity.
Source modeling determines the origin of the measured neuromagnetic fields, which is the goal of most MEG measurements. The mathematics behind this analysis is known as the inverse problem. Generally, the inverse solution of electromagnetic measurements is non-unique and ill-posed. If proper assumptions are made, however, the solution becomes solvable. We will discuss some of most commonly used source analysis methods used clinically.
In order to make the inverse solution tractable, one can approximate that the activity from primary sensory or motor cortex originates as a single equivalent current dipole (ECD). This is physiologically plausible, given that a limited patch of cortex is synchronously activated and that the sensors are at least a few centimeters from the source. The ECD provides spatial information, magnitude (current dipole moment), and direction. It is typically computed using a standard iterative least-square algorithm , which can also provide a measure of dipole parameter confidence, as well as the best-fitting parameters such as goodness of fit measure (GOF) .
Thus, by assuming a single source, the inverse problem has a unique solution. This works particularly well for primary sensory areas, focal epilepsies, and for higher cognitive areas that have a focal source. Further advances in the ECD approach, for both EEG and MEG, has made it possible to find multiple ECDs with a multidipole approach, such as that developed by Hari and colleges for somatosensory activation . One approach to investigate temporal changes in different areas of the brain is known as the time-varying dipole model. In this model, a series of dipoles are modeled such that the locations are fixed, but allowing the amplitudes to vary over time (See Figure 3). This ECD approach works quite well for sequentially or simultaneously activated cortical sources, although the fine spatial details are lost due to the fact that the measurements are obtained at least 3 cm from the brain sources .
If a large patch of cortex is activated, or if there are several areas of cortex activity simultaneously, the single ECD solution may be misleading. In practice this may be suspected when a dipole localizes too deep to be physiologically plausible, i.e., in the deep white matter. In such cases, distributed solutions such as the minimum norm estimate (MNE), the minimum current solution (MCE), or a beamformer solution maybe more accurate. Although numerous other inverse solutions exist [8, 9], this review concentrates on the MNE and beamformer solutions.
Originally pioneered by Hämäläinen , and improved upon by Dale and colleges [3, 4], the minimum norm estimate (MNE) is now available in various software packages, from freeware to commercially available software. The MNE does have some important limitations that must be kept in mind especially when used clinically: a depth bias, and a difficulty in determining the extent of activation. In its most elementary expression, the source variance is assumed to be equal throughout the volume, and the MNE solution is biased toward the most superficial currents. One approach to lessen this effect is to use a cortical constraint obtained from the anatomic MRIs . In addition to a depth bias, determining the extent of the sources is also problematic. Simulations  show that the point-spread function of estimates is a function of location. Further, the spread depends on the assumed source variance.
In order to compensate for the superficial bias of MNE current sources, Dale has suggested producing a dynamic statistical parametric map (dSPM) by normalizing the estimate MNE by the signal noise . Using the dSPM, the point-spread function is more uniform across the brain  and removes the superficial bias (See Figure 3). Additionally, the dSPM is F-distributed, allowing it to be used for hypothesis testing .
Another distributed source solution is commonly termed the beamform, which is based on the principal of spatial filters. Beamformers are typically are applied without a cortical constraint, although they can be modified to do so. The synthetic-aperture magnetometry (SAM)  is a beamformer approach that can be applied to both raw data and averaged (evoked) MEG or EEG. More sophisticated methods, such as Adaptive spatial filtering [14, 15] and Dynamic Imaging of Coherent Sources (DICS) , use a spectral analysis with coherence or other correlations which make it possible to measure functional connectivity in the brain within a particular spectral band. This makes it possible to localize changes in spectral power or functional connectivity that can be used to make evoked, such as language and motor mapping, or resting state functional connectivity.
In patients with epilepsy or brain masses, lateralization and localization of language functioning may be critical for preserving the quality of life. Currently, there are several methods to evaluate eloquent cortex, including neuropsychological testing, fMRI, MEG, the Intracarotid Amobarbital Procedure (IAP), and invasive electrode mapping [17–20]. Precise localization of the motor cortex and somatosensory cortex (central sulcus) is commonly required for frontal masses or for frontal lobe epilepsy. Verbal memory and language are the most common functions to be affected by a L-ATL, a common surgery for medial temporal lobe epilepsy [21–24]. Testing whether a temporal lobe is dominant for verbal function is often referred to as simply determining language lateralization (or verbal memory lateralization). Language cortex, typically the posterior language cortex, also needs to be precisely localized when a lesion is located in the posterior temporal lobe. We will briefly discuss the evaluation of language function in order to understand how MEG current evaluation of the surgical patient.
Neuropsychological testing assesses the functional status of the hippocampus and predicts the neurocognitive outcome of after resection of the hippocampus, but its use as a sole predictor of outcomes is limited. Neuropsychological test performance correlates with MTLE status and neuropsychological test performance before and after surgery, although the overall specificity is moderate [23, 24]. Material-specific memory loss as measured by neuropsychological test performance is observed in 20–30% of patients following unilateral ATL [24–26]. Material-specific memory loss is far more common following left- than right-sided ATL, with 25–50% of left anterior temporal lobectomy patients demonstrating verbal memory decline This is particularly true for patients without significant sclerosis on the surgical side. Declines in visual memory following right-sided ATL are less consistently observed but have been reported on tests that are highly spatial in nature. Accordingly, preoperative scores on material specific memory tests have demonstrated some utility in the prediction of cognitive morbidity following anterior temporal lobectomy; however, the sensitivity of prediction is modest (56%).
The Intracarotid Amobarbital Procedure (IAP) is used to predict language and memory outcomes following an ATL or for resections near cortical language area. For an IAP, amobarbital is injected into the internal carotid artery (ICA) through an arterial catheter and brief neuropsychological tests are performed. The IAP generally assesses verbal memory and, to a lesser extent, memory for visual items that are not easily encoded verbally such as scenes or faces [27, 28].
Despite its widespread use, some shortcomings to the IAP are widely known and limit its use [29–31]. It is invasive with a small but measurable risk of stroke, vascular damage, and infection [29, 31, 32]. The results of the IAP have been shown to be unreliable in test-retest studies, which may be due to flow of anesthetic to the opposite hemisphere (trans-hemispheric cross-flow) and variable penetration of the vasculature to the mesial temporal lobe from the internal carotid artery. Recently, there has been a shortage of amobarbital that has also reduced the use of the procedure, at least temporally . Pelletier, et al.  suggests four requirements needed before a test could replace the IAP: (1) A high predictive power for the presence or absence of critical functions in specific brain regions. (2) A user-independent statistical methodology. (3) High-spatial resolution (4) Production of reliable activation maps at the individual level. MEG meets these criteria and, in our experience, its routine use has lead to a reduction in the number of patients who undergo the IAP.
In part due to advancements in functional neuroimaging, the IAP is being used less and less frequently . Despite the IAP being the consensus gold standard for verbal functional, it has been criticized because of potential cross-flow to the contra-lateral hemisphere, and due to typical lack of testing of territory supplied by the posterior circulation. Since the mid 1990s, MEG is being used to determine both hemispheric dominance for language, as well as regional language mapping of individual language areas [35–43]. MEG of language areas provides an accurate non-invasive method of mapping language areas. In practice, MEG can be used to plan for an IAP and stimulation-based intraoperative mapping techniques.
Determining the location of language processing in a subject requires applying the task that best activates the desired stream of language processing. Specific language processes are complex and often include phonological, lexical, and also semantic processes. . Verbal memory encoding and retrieval occurs concurrently in virtually any task, making it difficult to separate `language processing' from `memory'. Further supporting language processes are attention, motor planning (speech), and visual or auditory functions. In fMRI, semantic decision and stem completion tasks are popular as they require a response from the patient, which ensures the quality of the patient's responses. For MEG, covert responses are more desirable, as overt (spoken) responses may lead to unacceptable motion artifacts. Still, some passive sensory paradigms requiring no patient response have been reported in the literature to be successful (see below).
Determination of the language dominant hemisphere is critical in the presurgical evaluation of patients who will have temporal or frontal lobe resections. For left anterior temporal lobectomies (L-ATL), the determination of language laterality is done routinely with MEG and fMRI. Wada and Rasmussen determined that over 93% of patients are left hemisphere dominant for language, as are over 96% of right-handed patients, although more recent studies indicate that many patients have more bilateral representation of language than the original studies . In left-handed epilepsy patients, only about 70% of patients demonstrate left hemispheric dominance for language, with about 15% of patients demonstrating bilateral language lateralization.
First proposed by Papanicolaou , sequential dipole fitting is a robust methods to determine language laterality (See Figure 5). The stimuli used by Papanicolaou were both auditory and visual words. The results concur with Wada test results  and electrical stimulation mapping . Szymanski et al  reported using simple phonetic stimuli, such as the vowel sounds /a/ and /u/, for determining the language hemispheric dominance by summing the number of selective dipoles in the late auditory magnetic field on each hemisphere, and calculating a lateralization index. Multiple groups report a strong correlation with both intraoperative mapping techniques and the results of the Wada test [39, 42].
In addition to lateralizing language, MEG also can accurately spatially locate language cortex. MEG localizes, with a high temporal resolution, both receptive and productive brain areas. The spatial resolution can be further increased by combining with a fMRI data . For localizing language specific areas in the cortex, there are several criteria that are important for patient studies compared with the more frequent use for cognitive neuroscience. First, the neurosurgical application requires precise localization in individual subjects, although for group studies the neuroscientist can average responses over several subjects. Second, presurgical evaluation usually requires mapping the essential language areas, not just participating areas. Removal of essential language areas may result in a language deficit. Participating areas are activated during language paradigms, but do not result in a post-operative language deficit after resection, either because there are areas of redundant processing or because other areas learn to take over the same functions. Currently, there is no reliable neuroimaging method to distinguish essential from participating areas with non-invasive imaging, and remains a major goal of clinical functional imaging. Combining MEG with other technologies such as transcranial magnetic stimulation (TMS) may make this possible in the near future.
By simply applying a source localization procedure, the same techniques described in determining hemispheric dominance can be used for regional language mapping. The mapping of equivalent current dipoles of the late auditory evoked fields can be used for both posterior temporal and frontal operculum mapping [46, 48] (Figure 5). Temporal maps of activation have similar profiles as determined by invasive electro-corticography (ECoG). On MEG tracing, the latency of Wernicke's area is typically between 210–420 msec, and Broca's area 400 msec – 700 msec, depending on the particular language task and differences in individual subjects. Generally, the peak activation of Wernicke's area precedes Broca's area, although occasionally other temporal profiles have been reported . The beamformer technique, such as SAM (Figure 6), can map decreases specific spontaneous activity power bands, such as beta band decreases (~20Hz), known as event related desynchronization .
A common clinical use of MEG is in the presurgical imaging of epilepsy or brain tumor patients is to aid in the localization of the central sulcus. Even with histologically low-grade tumors, the central sulcus can be distorted by tumor infiltration and/or mass effect, obscuring it on an anatomic MRI. Using a tactile stimulator, the somatosensory cortex homunculus can be easily mapped by successively stimulating finger digits, foot digits, and lip using MEG. Alternatively, an electrical nerve stimulator can be used to map the median, tibial nerve, and lip representative areas. If an electrical nerve stimulator is used, the electrodes are placed, and the intensity set, such that thumb twitching or toe twitching is elicited. After signal averaging, both the primary (SI) and secondary (SII) somatosensory cortices are detectable with MEG (Figure 3).
After recording the evoked magnetic fields, the primary somatosensory cortex is localized typically, such as with an ECD method [51–53] or distributed source [50, 54]. If electrical median nerve stimulation technique is used, the N20m—the first identifiable component of the evoked magnetic field—is easily evoked in most patients, including ones under deep anesthesia or in a coma [55, 56]. The N20m generator is located in the anterior wall of the somatosensory gyrus (Brodmann area 3b), with a non-radial orientation which is ideal for detection with MEG [7, 51, 53, 55, 57]. It is a pre-conscious field that does not synapse in the thalamus and, due to a high signal-to-noise ratio, has a very repeatable localization, with localization accuracy on the order of millimeters.
MEG identification of the central sulcus has been validated by several groups using intraoperative measurements using a variety of methods including equivalent current dipoles [45, 58–60] and the beamformer approach . Schiffbauer et al , for example, found that regardless of tumor grade, intra-axial brain tumors may border on or invade the somatosensory and auditory cortex. Importantly, low-grade tumors were more likely than high-grade tumors to invade and involve the functionally viable cortex. The authors concluded that low-grade tumors, due to slow growth, more often demonstrate functional activity within the radiologically abnormal areas, than high-grade tumors. High-grade tumors, on the other hand, often show functional activity at the margin of the contrast-enhancing area. They suggested these findings were caused by physical displacement of the functional tissue due to mass effect of the high-grade tumor. Firsching et al  reported that in 30 patients, ECD localization of the tactile neuromagnetic response localized in the somatosensory cortex, and was in agreement with phase reversal measurements at the time of surgery without exception. Further, they concluded that MEG-based functional neuronavigation was practical and reliable. Recent reports of non-invasive multimodal technologies—MEG, fMRI, and others—have noted that when combined enhance the reliability of identification of the central sulcus. Some have suggested using a functional risk profile (FRP), based on MEG findings to improve surgical decision-making . It should be kept in mind that brain tumor patients with known sensory or motor deficits may have diminished evoked fields . Nagaragan et al  demonstrated that an adaptive spatial filter, a type of beamformer, can accurately identify the motor cortex and the somatosensory cortex. They further demonstrated that it is be more accurate in some circumstances than the equivalent current dipole approach.
As noted above, many functional imaging techniques, such as fMRI and PET, can accurately identify the central sulcus . However, isolating pure motor activity for example, of the precentral gyrus with fMRI is difficult due to inevitable activation of the adjacent somatosensory cortex. Unlike fMRI and PET, MEG can be used to isolate motor activity due to high temporal resolution. In practice, it requires precise timing of the onset of motor movement in order to produce an averaged evoked field. This can be achieved by a self-paced button press, or by use of a trigger a photo-optic switch. Activity that peaks between 20–50ms before the onset of movement reflects activity in the primary motor cortex .
Alternatively, the motor cortex may be mapped by quantifying the functional connectivity between an electromyogram (EMG) and the MEG sensors, as suggested by Makela et al. . Placing bipolar electrodes over the first interosseous muscle and instructing the patient to slowly adduct generates large signal in both the EMG and MEG. Calculation of the coherence of the MEG-EMG yields a spectrum of a spike centered on or near 20Hz, strongest over the central sensors. This represents synchronization of the muscle twitches with the 20Hz `mu' rhythm, which is known to originate in the motor cortex . By localizing the functional connectivity, either by single ECD or distributed source method [xxxGross], identifies the primary motor cortex. This isolates the primary motor cortex from the somatomotor network, something that can be difficult with fMRI. Further, MEG identifies the entire neural network activated during the planning and the act of motor movement, including supplementary cortex [66, 67] and premotor cortex [16, 68].
In patients with lesions lying near visual areas, MEG can evaluate the function of visual cortex. Mapping of the visual areas is theoretically difficult withm MEG due to synchronously active sources that may result in magnetic field cancellation at the scalp. The primary visual cortex can be mapped by a simple ECD of the first visual evoked field from nearly any strong visual stimulus, typically a portion of a visual quadrant is stimulated with a checkerboard. In practice, however, it does yield valid results [69–71]. Mapping of the magnetic equivalent of visual evoked potentials N75, P100, and N145 components is robust with large visual stimuli with phase reversal techniques, and can be performed to detect visual field deficits [69–71].
Localization of epileptic discharges may be performed in the context of a low-grade brain tumor or as part of the workup of surgical epilepsy. Interictal activity is easily captured with MEG. Ictal activity, although more difficult, can also be detected and localized with whole head MEG systems—particularly if the instrument is located in a hospital where antiepileptic medicines can be tapered. The most effective MEG measurement of epilepsy is with a whole-head system with simultaneously recorded EEG. A standard EEG electrode array, manufactured with no magnetic material, provides a standard EEG classification of ictal or interictal activity. High-density electrode caps are also available if EEG source localization is needed.
Presurgical evaluation of patients with epilepsy typically involves simultaneous whole-head MEG and EEG recording during rest, which records basic rhythmic activity, sleep-related activity, and other spontaneous brain activity, in addition to the epileptic discharges. Localizing the neural source of interictal spikes present in the MEG data (see Figure 7) localizes the epileptogenic zone even in the event of propagation [72–75] The equivalent current dipole (ECD) is an important source model for epileptic discharges, typically for localizing the peak of an ictal or inter-ictal epileptic discharge. However, the ECD at times may not match the location of the seizure onset. In particular, source analysis using a distributed source model, such as minimum norm estimate (MNE) , and beamformer techniques  may be helpful in demonstrating the time-course of cortical activation and thereby reveal the generation and propagation of epileptic activity .
A number of recent studies demonstrate that MEG improves the quality clinical care in refractory epilepsy patients. Including the MEG in the presurgical evaluation of epilepsy influences clinical decisions and increases the likelihood of surgical success (making a person seizure free) [79–81]. MEG also provides non-redundant clinical information in about one third of the epilepsy cases that were performed during the presurgical evaluation .
Functional connectivity defines correlations of activity in regions across brain regions, whereas effective connectivity measures the influence of brain regions on each other. Corticomuscular coherence is a type of functional connectivity between the motor cortex and the muscle subunits that is used clinically to identify primary motor cortex. Functional connectivity using temporal correlations in fMRI has made it possible to map several neural networks, such as the motor system . The high temporal resolution of MEG allows connectivity analysis to be performed at a high temporal resolution. Other abnormalities in functional connectivity include epilepsy. For example, abnormally high functional or effective connectivity across small distances could influence epileptic spike generation. Increased connectivity across brain areas has been hypothesized to be responsible for the propagation and generalization of epilepsy. On the other hand, a failure of integration of perceptions might suggest a failure in “binding” object features within and across brain regions, may lead to psychiatric diseases such as schizophrenia. Functional connectivity, using MEG may help define epileptogenic cortex, especially in multifocal epilepsy .
Recent work has shown that abnormalities in brain connectivity as measured by MEG have clinical value, such as in patients with brain tumors . Patients with brain tumors have abnormal connectivity compared to healthy controls (Figure 8). In patients with brain tumors, eloquent cortex found within abnormal resting-state functional connectivity have an increased risk of post-operative deficits (Figure 9). These are some potentially new useful clinical applications of MEG in neurosurgical patients.
Co-registration of MEG and structural images with skin fiducial markers makes it makes it possible to delimit the volume containing the lesion and the surgical field [85–87]. Frameless stereotactic systems assist with resection of small, deep-seated tumors during neurosurgery, with a precision approaching 2mm. Ganslandt et al  combined MEG somatosensory mapping with a free-hand stereotactic pointing device in 25 cases of peri-rolandic tumors and masses. They found agreement with intraoperative somatosensory mapping in all cases. Further, they eliminated 11 patients who were not available for open surgery because of tumor infiltration of the motor cortex . Rezai et al [89, 90] report the use of a combination of MEG-derived functional mapping, and CT scans, MRI, and digital angiography reduces surgical risk. In order to overcome the errors associated with shift of brain contents during an operation requires sulcal landmarks on 3-D reconstructions with fMRI, MEG, iEEG and tractography aid in surgical planning.
MEG is now an accepted method of presurgical evaluation in patients with brain tumors and epilepsy. Beyond epileptic discharge mapping and non-invasive mapping of eloquent cortex, clinical research studies suggest that MEG is ready for other neurosurgical applications, including head trauma, Parkinson's disease, and other disorders. Combining MEG, EEG, fMRI and new methods of connectivity are also being evaluated for clinical applications. These studies will continue to change how MEG is applied in neurosurgical evaluations.
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.