The term magnetoencephalography (MEG) refers to the recording of the magnetic field of electrocerebral activity. MEG uses specialized equipment and lends itself to source localization using dipole methods and co-registration with magnetic resonance imaging (MRI). Because of the properties of the magnetic field and its lack of interaction with biologic tissues at the frequencies of brain electric currents, the MEG is inherently more accurate and less subject to distortion than is the electroencephalogram (EEG). The co-registration of MEG and MRI is called magnetic source imaging (MSI).
The MEG is generated by the small electric currents set up in neurons during postsynaptic potentials. Because of the symmetry of the magnetic field, MEG records primarily intracellular currents oriented tangentially or parallel to the scalp surface, whereas EEG records extracellular volume currents produced by neurons of all orientations. Figure 10-1 shows the MEG and EEG field patterns expected for a tangential source and for two radial sources.
As with the EEG, summation of the activity of a large number of parallel-oriented neurons is necessary to produce a detectable signal outside the head. These conditions are met by the larger pyramidal neurons in cortical layers III to V. The MEG is generated according to the right-hand rule of physics, by which current in a small segment of wire flowing from positive to negative in the direction of the thumb of the right hand will generate a magnetic field in the direction of the curled fingers. The magnetic field map or pattern recorded over the scalp is generated by the electrical current in a small cortical region. This map shows a point of maximum, exiting field-strength or peak; a point of minimum, entering field-strength or valley; and a null point between the peak and valley. This is a “tangential dipolar” pattern and is the usual appearance of the MEG isocontours. This pattern is invariant and easy to model.
To record the MEG, two problems had to be solved. First, a sensor with high sensitivity was required to record the very small signal of interest. The MEG is less than one picoTesla (10 -12 Tesla), or about a billion times smaller than the earth’s magnetic field (10 -4 Tesla). Second, the magnetic fields in the environment create a virtual magnetic storm, and thus a signal-to-noise problem. Magnets used for MRI (about 1 Tesla), fluorescent lights, and cars moving in the street all produce magnetic signals that are much larger than the MEG. This situation is analogous to trying to measure the fluctuating height of a microscopic ripple on a pond while the pond is in the path of a hurricane.
The solutions were ingenious. The problem of sensitivity was solved by the use of a superconducting quantum interference device (SQUID) immersed in liquid helium (at 4° Kelvin, or 4° above absolute zero). Interference devices are very sensitive instruments to measure waves in a wide range of applications. A SQUID is a superconducting ring of metal, with one conducting wire leading current into the ring and another leading current away from it. The current passes through the metal of the ring by splitting into two halves of current in each of the two semicircles of the ring and then rejoining to flow out of the ring into the exiting conductor wire. There is a “weak link” or segment where the metal is pinched down to a smaller diameter in each of the two parts of the ring, so that much less current can pass through the ring. Each of these weak links has an upper limit of current that can pass through it, called a critical current . While the ring is supercooled in liquid helium and superconducting, there is no resistance in the wire. When, however, a change in magnetic field passes through the center hole of the ring, the magnetic field induces a circular current in the ring according to Lenz’s law. This increases the current flowing through one of the weak links to above the limit of the critical current. This stops the ring from superconducting, and the ring then has much higher resistance and conducts much less current. There is a very large change in the current passing through the ring in going from a superconducting state, in which there is zero resistance and a very large current, to a normal state, in which there is a large resistance and much smaller current. This change in current is measured easily at a macroscopic level, so that the MEG, although very small, induces a large change in current in the SQUID.
Two innovations helped to solve the second problem of the signal-to-noise ratio. First, the development of a magnetically shielded room canceled out all environmental noise. Second, a gradiometer coil enabled the magnetic fields from the heart to be canceled out and the magnetic fields to be recorded selectively near the head directly under the sensor. The gradiometer is still needed inside the shielded room to cancel out the heart’s magnetic field (MKG), which is one hundred times larger than the MEG. Using a SQUID attached to a gradiometer coil inside a shielded room allows recording of MEG, even in the hostile noise environment of a modern imaging center or hospital.
Figure 10-2 shows a subject having an MEG study under the dewar of a whole-cortex neuromagnetometer inside a magnetically shielded room.
Although MEG recording equipment is complex and costly, the MEG, once obtained, is easier to analyze than the EEG because of its simpler geometry and invariant field patterns. The first step is reviewing the raw time-series data. This is almost the same as the EEG; the MEG shows alpha rhythm, sleep spindles, evoked responses, and abnormal activity such as spikes. The second step is reviewing the isocontour maps over the head at a single time-point. A map with complementary extrema (maximum and minimum) and a null line between usually indicates an underlying source, which can be modeled as a tangential dipole.
The basic idea of localization by a single dipole model is straightforward. First, a computer program performs a simulation. It places a dipole or infinitesimally small segment of wire with current at a specific location and orientation inside the skull. Then the magnetic field pattern generated by the dipole is compared with the measured MEG field pattern. A goodness-of-fit is calculated using the squared difference between the computed field and the measured field. For that location and orientation, the goodness-of-fit is placed in a table. Then the computer moves the dipole to a new location and new orientation and repeats the whole process, storing the results on the next line of the table. This is repeated many times until the computer finds the location and orientation that has the least squared difference or smallest error between the computed and measured field. The computer then reports that location and orientation as the true location and orientation of the center of the brain electrical activity.
If there were no noise, and if all assumptions of the model were correct, the computed location and orientation would coincide with the actual parameters of the brain activity. Unfortunately, in the real world, there are noises and complexity in the measurements that prevent perfect estimates. The sources of error can be broken down into the parts of the model.
Because of its simpler nature, the MEG lends itself to thorough analysis of the different contributions of noise to its modeling errors. The contributions to localization error arise from several parts of the model, which are best visualized going from the brain to the recording at the SQUID. First, the source of current must be modeled correctly as a single focal region, as multiple regions, or as an extended layer of brain activity. Second, the background activity of the brain or “biologic noise” must be modeled to separate it from the activity of interest, such as an epileptic spike or evoked response. Third, the volume conductor of the head through which the magnetic field travels to get to the sensor must be modeled. Fourth, the sensors must be co-registered to the MRI in correct relation to the brain structures recorded. Fifth, the instrumentation noise of the measurement system, including the inherent noise of the SQUID and the SQUID electronics, must be modeled.
The source model chosen for clinical applications usually is the single dipole discussed earlier. Any type of brain activity usually involves a large ensemble of neurons in a cortical layer of finite extent. However, when the recording sensor is far enough away from a single region of brain activity, the source can be modeled as a point source at the center of the cortical activity. In other applications, the use of two or more dipoles is necessary because there are multiple regions that are simultaneously active. In still other models, no attempt is made to assume a specific type of source, such as a dipole; rather, an estimate is made of the distribution of currents circulating in a wide region. Each of these models has its own advantages and disadvantages; however, the single dipole is useful for most clinical applications.
The signal-to-noise ratio can be improved to detect smaller sources (e.g., sharper initial spikes) using advanced spatiotemporal models such as the recursively applied multiple signal classification algorithm (designated by the acronym RAP-MUSIC), which separates the signal from the noise subspace, increasing the signal-to-noise ratio. More realistic modeling may improve detection sensitivity of MEG. More complex modeling has been used in scientific work by several investigators. Propagation is assessed best with the more realistic spatiotemporal dipole model.
The volume conductor for MEG can be modeled simply as a single sphere or local sphere that is homogeneously conducting. Straightforward equations exist that are programmed easily for this application. The EEG, by contrast, arises from volume currents, and the tissue interfaces and conductivities of different brain, cerebrospinal fluid, skull, and scalp layers must be modeled. This is a more difficult problem because these conductivities are not known a priori and are quite variable from patient to patient.
The model of the volume conductor significantly affects localization accuracy of dipole and extended source models. Several mathematical and geometric models of the human head as a volume conductor have been developed to correlate extracranially measured magnetic fields (i.e., by MEG) and electrical potentials (i.e., by EEG) to their intracranial generators. Models such as multiple concentric spherical shells, the boundary element model, and the finite element model are among the most widely used. The boundary element model divides the volume conductor into closed surfaces. The finite element model divides the volume conductor into many small tetrahedrons; it is more realistic but is also computationally intensive. Since exact conductivities are not known in the individual patient, it is also subject to misspecification errors. The boundary element model is a compromise. Skull effects usually dominate in the more complex volume conductor models. Detailed studies of conductivity have revealed the complexity of the skull’s electrical properties. Inclusion of the results of such studies in the volume conductor model eventually may improve localization from extracranial electrical and magnetic measurements.
The co-registration of sensor sites with the MRI is easier for the MEG than the EEG. The MEG sensors are in a fixed array inside the dewar flask containing the liquid helium. Thus, only the location and orientation of the array center and axis need to be measured, rather than calculating all the locations of each of the electrodes attached to the nonspherical surface of the scalp, as is required for the EEG. This reduces the errors in MEG.
Although, theoretically, EEG is less accurate than MEG for spatial localization, and EEG sensors are more difficult to co-register with MRI, EEG has several complementary advantages. It is used routinely in all comprehensive epilepsy centers in the United States. It is less expensive than MEG because it records only the voltage differences between two sites on the head, even though using the recording of this easily obtainable quantity increases the complexity of the volume conductor necessary to analyze it. Recording of epileptic activity usually is easier by EEG, which can be used for long-term studies. The electrodes are attached to the scalp, and the patient can move around while being recorded. This allows the recording of seizures and a large spike sample across wake–sleep states to assess autonomy. The portability of EEG allows recording in the operating room and intensive care unit. EEG measures the same current orientations as intracranial electrographic recordings. EEG also measures more of the cerebral cortex than MEG, concurrently increasing the complexity of the source model required for analysis. The EEG combined with the MEG allows more accurate localization than either MEG or EEG alone. More realistic volume conductor models are especially important to improve the accuracy of the EEG and are essential to combine the MEG and EEG for localization purposes.
Validation studies of localization accuracy
MEG localization is accurate. The results from the author’s laboratory are reported here as representative of the accumulating literature. The localization error in MEG was quantified for a dry dipole phantom for one dipole and two dipoles. It was found that the MEG method, analyzed with realistic modeling, gives accurate localizations for one dipole and for two dipoles overlapping in time.
The dry phantom consisted of two flat plastic semicircular plates fitted together at a right angle along their diameters to mimic a sphere. Small wire current dipoles were embedded in the plastic, tangential to the radii. This is used routinely as a calibration phantom. A single dipole source and a two-dipole source were tested at multiple locations using RAP-MUSIC from National Institutes of Health-funded BrainStorm software, developed as shareware for scientists. The phantom was attached mechanically to the bottom of the dewar inside the space for the head in a 100-SQUID 68-sensor channel whole-cortex neuromagnetometer. The neuromagnetometer used 68 sensor first-derivative co-axial gradiometers and 19 reference channels to produce synthetic third-derivative gradiometers, which were flux-transformed to direct-current SQUIDS. Coil parameters were baseline 5 cm and diameter 2 cm. Channel noise was 5 to 7 femtoTesla per root Hz.
For one-dipole activation, mean localization error was 1.0 mm (SEM, 0.4), with 95 percent confidence that the actual location was less than 2 mm from the estimated location. For two dipoles, overlapping in time but offset by 100 msec, mean error was 1.7 mm (SEM, 0.6) for each, with 95 percent confidence that each of the dipoles were within 3 mm of the estimate.
With perfectly synchronous activations of identical waveforms, errors were larger. Such synchrony may not be relevant to human cortex because of delays from finite conduction speed of cortico-cortical association U-fibers. Even in 3-Hz generalized spike-wave activity in primary generalized absence epilepsy (classic “petit mal”), which is the prototype of generalized epileptiform activity, interhemispheric conduction delays are 15 to 20 msec.
Source configurations with two dipoles on the same radius also gave larger errors. This problem, anticipated theoretically, has been approached historically in at least two ways. First, experimental design can help. For instance, the auditory P300 has components from superficial and deep structures that overlap in time and space (e.g., the primary auditory cortex, hippocampus, and lateral neocortex). These structures are in the same temporal lobe and, during some parts of the activity, on similar radii. Investigators have addressed this problem by studying the visual P300, where the primary cortical response is in the occipital lobe and the later response is in the medial temporal lobe. In this case, the fields of the earlier and later responses have less overlap. Thus innovative experimental design helps avoid situations in which larger localization errors result from co-radial sources.
Additional studies showed only slightly larger mean errors of localization when the experiments were repeated in a saline-filled sphere phantom with one dipole. In that case, the mean localization error was 1.4 mm and the SEM was 0.9 mm.
Combined MEG and EEG methods may help. For instance, comparison of the timing of an EEG component at the deeper sphenoidal electrode with the timing of a component at a lateral scalp electrode may help to distinguish deep and superficial sources. The additional sensitivity of an EEG electrode contact recording extracranially but directly under the foramen ovale thus can add complementary sensitivity and spatially specific information to address the problem of temporally overlapping sources that are superficial and deep, with one under the other.
Median Somatosensory Evoked Field
The localizations of the median somatosensory evoked response (SER) have been tested on whole-head neuromagnetometers and replicated between a 68-sensor system (with recordings made in 1998) and a 150-sensor system (localizations made in 2001). The two laboratories were in different cities, about 2,000 miles apart. The electrical stimuli were delivered to the median nerve at the wrist to give a reliable thumb twitch at 3.1 Hz, to avoid 60-Hz line-frequency. The parameters were digitization rate 1,250 Hz, bandpass 0.1 Hz to 300 Hz, and 250-msec sweep with a 50-msec prestimulus baseline, with a subsequent high-pass filter to eliminate low frequencies and later activity, which were not of interest in this study. Localizations from the single moving dipole model were applied to the SER. The latencies of the best-fitting dipoles were slightly different in the two studies but within experimental error for the more stable P30m (P1m). The localization from each laboratory was co-registered on the subject’s MRI.
The mean euclidean distance between the co-registered localizations of the two systems was 2.1 mm (standard deviation [SD], 1.84), with the largest error vertically, on the z-axis. This difference is within 1 SD of the localization error in the saline-filled sphere phantom, and similar to the 2- to 3-mm error in a phantom study on a 122-channel whole-head system. Thus, within the error limits derived from phantom studies, there was no detectable difference in localization of the early dipolar components of the median SER measured on two different neuromagnetometers many miles apart and over a period of 3 years. This confirms the theoretical predictions of MEG localization accuracy and the lack of a significant difference between 64 and 128 channels for dipole localization of cortical sources.
The comparison of the two studies was facilitated by the ease of data transfer and superimposition in the same programs on the same co-registered MRI. Accurate localization of the superficial focal currents producing the early components of the SER is one of the most difficult tests of a localization method. A superficial source has the highest spatial frequency of topographic map isocontours.
Comparison of Seizure Discharges
The clinical procedure of afterdischarge mapping to determine neocortical epileptogenicity has been used in another test of localization accuracy. Simultaneous MEG and electrocorticogram (ECoG) was recorded for an electrically induced focal afterdischarge seizure in the frontal lobe of a patient undergoing evaluation for epilepsy surgery. In this case the origin of the seizure was known to be in the cortex directly under the stimulating electrode. Dipole localization was applied to the repetitive spiking of the seizure in the ECoG and MEG measured at the same time.
Figure 10-3 shows that the MEG and ECoG localizations were within about 1 cm of the center of the subdural ECoG electrode where the afterdischarge was induced. There was no significant difference between the MEG error (mean 8.5 mm, SEM 1.5) and the ECoG error (mean 10.4 mm, SEM 0.3). In this study, the MEG and ECoG were both able to record the same event with a similar signal-to-noise ratio, and their localizations were indistinguishable. The ECoG usually has been regarded as superior to noninvasive electrophysiologic methods in both specificity and sensitivity. This study suggests, however, that with similar sensitivities the MEG and ECoG can have equivalent localization specificity for pathologic activity that is the definition of epilepsy (i.e., the seizure itself). This finding is important in suggesting a means to improve noninvasive localization of abnormal activity for focal excisional surgery in epilepsy centers. Many patients now require invasive recording with ECoG or depth electrodes to localize and quantify the zone of seizure origin. This study suggested that MEG may provide useful preoperative information to guide electrode placement to avoid sampling error such as edge-of-the-grid effect.