Imaging the Human Brain with Magnetoencephalography

abstract

Magnetoencephalography is a relatively new medical imaging modality for the monitoring and imaging of human brain function. Extracranial magnetic fields produced by the working human brain are measured by extremely sensitive superconducting sensors, called SQUIDs, enclosed in a liquid helium-filled dewar. Mathematical modeling allows the formation of images or maps of cortical neuronal currents that reveal neural electrical activity, identify cortical communication networks, and facilitate the treatment of neuronal disorders, such as epilepsy.

introduction

Magnetoencephalography (MEG) is a noninvasive technique for measuring neuronal activity in the human brain. Electrical currents flowing through neurons generate weak magnetic fields recorded using magnetic sensors surrounding the head. The MEG method is part of a broad area of research referred to as biomagnetism, which involves studies of magnetic fields emanating from several organs of the human body, notably the brain and heart.

The temporal resolution of MEG is in the millisecond (ms) range, the timescale at which neurons communicate. Therefore, we can follow the rapid cortical activity reflecting ongoing signaling between different brain areas. This is a great advantage compared to other medical imaging modalities such as functional magnetic resonance imaging (fMRI) and positron emission tomography (PET), where temporal resolution is on the order of seconds. Furthermore, unlike other methodologies that measure brain metabolism or the relatively slow hemodynamic response, MEG directly measures electrical brain activity. Electroencephalography (EEG) is a complimentary method to MEG, measuring electrical scalp potentials rather than magnetic fields. It offers similar temporal resolution to MEG, but the spatial resolution is less accurate because electrical potentials measured on the scalp are heavily influenced by strongly inhomogeneous conductivity of the head, whereas magnetic fields are mainly produced by currents that flow in the relatively homogeneous intracranial space.

NEURAL BASIS OF ELECTROMAGNETIC SIGNALS

A neuron consists of the cell body (or soma), which contains the nucleus; branching dendrites, which receive signals from other neurons; and a projection called an axon, which conducts the nerve signal. When a pulse arrives at an axon of a presynaptic cell, neurotransmitter molecules are released from the synaptic vesicles into the synaptic cleft. These molecules bind to receptors located on target cells, opening ion channels (mostly Na+, K+, and Cl-) through the membrane. The resulting flow of charge causes an electrical current along the interior of the postsynaptic cell, changing the postsynaptic potential (PSP). When an excitatory PSP reaches the firing threshold at the axon hillock, it initiates an action potential that travels along the axon with undiminished amplitude.

The conservation of electric charge dictates that intracellular currents, commonly called primary currents, give rise to extracellular currents flowing through the volume conductor. Both primary and volume currents contribute to magnetic fields outside the head; however, only locally structured arrangements of cells can achieve sufficient coherent superposition of currents as to produce measurable external fields. Clusters of thousands of synchronously activated pyramidal cortical neurons are believed to be the main generators of MEG signals (Figure 1). In particular, the currents associated with large dendritic trunks, which are locally oriented in parallel and perpendicular to the cortical surface, are believed to be the primary source of the neuromagnetic fields outside the head. In contrast, the temporal summation of currents for action potentials, which have duration of only 1 ms, is not as effective as for dendritic currents flowing in neighboring fibers, so action potentials are believed to contribute little to MEG measurements.

Figure 1. Cerebralfrontal cortex drawn by Ramon Cajal using a Golgi staining technique. Pyramidal (A, B, C, D, E) and nonpyramidal (F, K) cells are clearly depicted. Currents flowing in the dendritic trunks of pyramidal cells are believed to be the primary generators of magnetic signals outside the head.

Cerebralfrontal cortex drawn by Ramon Cajal using a Golgi staining technique. Pyramidal (A, B, C, D, E) and nonpyramidal (F, K) cells are clearly depicted. Currents flowing in the dendritic trunks of pyramidal cells are believed to be the primary generators of magnetic signals outside the head.

INSTRUMENTATION

Empirical observations indicate that we observe sources on the order of 10 nA-m, and consequently, the neuromagnetic signals are typically 50 to 500 fT, that is, 109 or 108 times smaller that the geomagnetic field ofthe earth (Hamalainen, Hari, Ilmoniemi, Knuutila, & Lounasmaa, 1993). The only detector that offers sufficient sensitivity to measure such fields is the superconducting quantum interference device (SQUID) introduced in the late 1960s by James Zimmerman (Zimmerman, Thiene, & Harding, 1970). The first measurement of brain magnetic fields using a SQUID magnetometer was carried out by David Cohen (1972) at the Massachusetts Institute of Technology, and it consisted of the spontaneous alpha activity of a healthy participant and abnormal brain activity in an epileptic patient.

The SQUID is a superconducting ring interrupted by thin insulating layers to form one or two son junctions (Barone & Paterno, 1982). One important property associated with son j unctions is that magnetic flux is quantized in units of . If a constant biasing current is maintained in the SQUID device, the measured voltage oscillates as the magnetic flux increases; one period of voltage variation corresponds to an increase of one flux quantum. Counting the oscillations allows one to evaluate the flux change that has occurred, and therefore detect magnetic fields on the order of a few fT. The sensitivity of the SQUID can be increased to 1 fT by attaching a coil of superconducting wire or flux transformer. The latter is placed as close to the human head as possible, and depending on its shape, it can be configured as a first-order planar or axial gradi-ometer, a second-order axial gradiometer, or a simple magnetometer (Hamalainen et al., 1993). The gradiometer configurations produce measurements proportional to the spatial gradient of the magnetic field, thus offering robustness to interference from distant magnetic field sources.

Modern MEG systems consist of a few hundred SQUID sensors placed in a liquid-helium-filled dewar, with the flux-transformer pickup coils surrounding a helmet structure (Figure 2). Worldwide, three companies build the majority of whole-head MEG systems: 4-D Neuroimaging (formerly Biomagnetic Technologies Bti), Elekta Neuromag Oy, and VSM MedTech Ltd. (manufacturers of the CTF Systems). In recent years, all three vendors have introduced dense arrays comprising over 200 SQUID channels.

Brain magnetic signals are very weak compared to ambient noise. Outside disturbances include fluctuations of the earth’s geomagnetic field, power-line fields, electronic devices, elevators, and radio-frequency waves. Nearby artifacts are caused by instrumentation noise and body interference, such as heart, skeletal muscle, and spontaneous or incoherent background brain activity. Shielded rooms made of successive layers of mu-metal, copper, and aluminum effectively attenuate high-frequency disturbances. Furthermore, gradiometer flux transformers cancel distant noise sources that produce magnetic fields with small spatial gradients.

Figure 2. Whole-head CTF Omega MEG system with 275 axial gradiometers (left), and MEG sensors using low-temperature electronics cooled by liquid helium (right)

Whole-head CTF Omega MEG system with 275 axial gradiometers (left), and MEG sensors using low-temperature electronics cooled by liquid helium (right)  

MODELING

To estimate the neural sources of magnetic fields, one must first solve the associated forward problem, that is, the forward model that maps sources of known location, strength, and orientation to the MEG sensors. The most common source model is the current dipole (Baillet, Mosher, & Leahy, 2001), used to approximate the flow of an electrical current in a small area of the brain. The typical strength of a current dipole, generated by the synchronous firing of thousands of neurons, is 10 nA-m. Alternatively, to avoid the identifi-ability problem that arises when too many small regions and their dipoles are required to represent a single large region of coherent activation, we can use multipolar models, consisting of dipoles, quadrupoles, octupoles, and so on (Mosher, Leahy, Shattuck, & Baillet, 1999).

Since the useful frequency spectrum for electrophysiological signals is largely below 100 Hz, the physics of MEG can be described with the quasistatic approximation of Maxwell’s equations. The propagation of electromagnetic fields inside the head is estimated based on the conductivities of the scalp, skull, gray and white matter, cerebrospinal fluid, and other tissue types. Head models that consist of a set of nested concentric spheres with isotropic and homogeneous conductivities have closed-form solutions. Even though spherical head models work surprisingly well, more accurate solutions use realistic head models based on anatomical information from high-resolution magnetic resonance (MR) or x-ray computed tomography (CT) volumetric images. To estimate the parameters of these models, numerical solutions using boundary-element methods (BEMs), finite-element methods (FEMs), or finite-difference methods (FDMs) are necessary (Darvas, Pantazis, Kucukaltun-Yildirim, & Leahy, 2004).

To make inferences about the brain activity that gives rise to a set of MEG data, we must solve the inverse problem, that is, find a neuronal current-source configuration that explains the MEG measurements. Inverse methods for MEG can be roughly categorized into two classes: imaging methods and dipole-fitting or -scanning methods. The imaging approaches are based on the assumption that the primary sources are intracellular currents in the dendritic trunks of cortical pyramidal neurons that are aligned normally to the cortical surface. Consequently, a tessellated representation of the cerebral cortex is extracted from a coregistered MR image, and the inverse problem is solved for a current dipole located at each vertex of the surface. In this case, since the position and orientation of the dipoles are fixed, image reconstruction is a linear problem and can be solved using standard techniques. The dipole-fitting or -scanning methods assume that the sources consist of only a few activated regions, each of which can be represented by an equivalent current dipole of unknown location and orientation. The standard approach to localization is to perform a least-squares fit of the dipole model to the data (Lu & Kaufman, 2003). More recently, scanning methods have been developed that are also based on the dipole model, but involve scanning a source volume or surface and detecting sources at those positions at which the scan metric produces a local peak (Baillet et al., 2001). Examples of these methods include the MUSIC (multiple signal classification) algorithm (Mosher, Leahy, & Lewis, 1992) and the LCMV (linearly constrained minimum variance) beamformer (VanVeen, van Drongelen, Yuchtman, & Suzuki, 1997).

Figure 3. MEG model depicting: (a) Sensor arrangement of a275-channel CTF MEG system, (b) topography of sensor measurements, and (c) minimum-norm inverse solution on a tessellated cortical surface

MEG model depicting: (a) Sensor arrangement of a275-channel CTF MEG system, (b) topography ofsensor measurements, and (c) minimum-norm inverse solution on a tessellated cortical surface

Due to intrinsic spatial ambiguities of the electromagnetic principles that underlie MEG, the spatial resolution is lower than that of PET and fMRI. These ambiguities force a choice between low-resolution linear cortical imaging methods, or potentially higher resolution methods based on parametric models, or Bayesian or other nonlinear imaging methods incorporating physiological priors that reflect the expected characteristics of neural activation. A consensus is developing in the research community that no single method suits all MEG applications; each method has strengths and weaknesses, reflecting the ill pose of the inverse problem. The characteristics of expected neural activation, as well as model-fitting techniques, can facilitate the proper choice of inverse methodology.

statistical analysis

Given the large number of localization methodologies, it is important to perform validation and statistical analysis under different experimental settings, such as the number, location, and time series of neuronal sources. Furthermore, several methods require the fine-tuning of parameters, such as the subspace correlation threshold for the MUSIC algorithm. The receiver operating characteristic (ROC) curve is a standard tool to evaluate the trade-off between sensitivity and specificity, and to compare different inverse methods. By varying a threshold applied to localization maps, we can estimate two performance measures: the sensitivity or true positive fraction (TPF), and 1-specificity or false positive fraction (FPF). The ROC curve is a plot of the TPF vs. FPF as a detection threshold is varied. When comparing two detection methods, the one whose ROC curve gives higher sensitivity at matched specificity, and vice versa, for all points on the curve is the better detector. A simple metric to compare methods is the area under the ROC curve (AUC), where the method with the largest AUC is superior. The use of free-response ROC, an ROC variant that can handle the presentation and detection of multiple targets per image, is demonstrated in Yildirim, Pantazis, and Leahy (in press) for the evaluation of minimum-norm and scanning-inverse methods.

In addition to evaluating the relative performance of different methods, it is important to establish some degree of confidence in the results of real data analysis. Dipole-scanning methods often produce unstable solutions, and the reproducibility of the reconstructed dipoles is not guaranteed. A number of different approaches have been investigated for assessing dipole-localization accuracy, including Cramer Rao lower bounds, perturbation analysis, and Monte Carlo simulation. To avoid strict distributional assumptions, a resampling alternative based on bootstrap theory was proposed in Darvas, Rautiainen, et al. (2005). The principle underlying the bootstrap theory is that although the distribution of the data is unknown, it can be approximated by the empirical distribution of a set of independent trials. By sampling with replacement over independent trials collected during an event-related MEG study, the position, variance, and time series of current dipoles can be estimated reliably.

In contrast to dipole-scanning methods, imaging methods are hugely underdetermined, resulting in low-resolution localization maps; interpretation is further confounded by the presence of additive noise exhibiting a highly nonuni-form spatial correlation. In this case, we need a mechanism to decide which features in the data are indicative of true activation vs. those that are noise artifacts. To determine a suitable threshold for detecting statistically significant activation, the familywise error rate (FWER), that is, the chance of any false positives under the null hypothesis of no activation (Type 1 error), is typically controlled. Parametric random-field methods and nonpara-metric permutation methods are used to estimate familywise-corrected thresholds in Pantazis, Nichols, Baillet, and Leahy (2005). Alternatively, the control of the false discovery rate (FDR), that is, the proportion of false positives among those tests for which the null hypothesis is rejected, can produce more sensitive thresholds.

Recent literature in MEG statistical analysis has been mostly limited to pairwise comparisons at each cortical surface element for event-related averages. However, extensions of this methodology to the investigation of multiple effects using analysis of variance (ANOVA) and analysis of covariance (ANCOVA) in individuals and groups is possible, as, for example, in Brookes et al. (2004).

applications

Applications in MEG include both basic and clinical research. One of the most important clinical applications is the detection, classification, and localization of abnormal neuronal activity in epilepsy patients. MEG has been successfully used to localize three different spontaneous interrictal signal components: epileptic spikes, slow-wave activity, and fast-wave activity (Lu & Kaufman, 2003). The neurosurgical planning of medically intractable epilepsy often includes the identification of epileptogenic lesions with MEG (Ossadtchi, Baillet, Mosher, Thyerlei, Sutherling, & Leahy, 2004; Stefan et al., 2003). Furthermore, recent literature investigates the possibility of seizure prediction based on a drop in the complexity of neural activity immediately before seizures (Maiwald, Winterhalder, Aschenbrenner-Scheibe, Voss, Schulze-Bonhage, & Timmer, 2004).

In addition to the diagnosis of epilepsy, MEG is currently used for functional brain mapping. Evoked response fields have been used to identify somatosensory-, motor-, and vision-related activity (Lu & Kaufman, 2003). Several MEG studies (Pantazis, Merrifield, Darvas, Suther-ling, & Leahy, 2005) have localized language-specific cortical activity using either equivalent current dipoles or distributed cortical imaging, with promising results for clinical application in neurosurgery. Time-frequency analysis of MEG oscillatory-evoked responses (Pantazis, Weber, Dale, Nichols, Simpson, & Leahy, 2005) has detected networks of cortical interactions and determined the functional specificity of several frequency bands. A wide range of signal-processing techniques including image modeling and reconstruction, blind source separation, phase synchrony estimation, nonlinear analysis, and chaos theory are under investigation to reveal complex cognitive processes such as attention and working memory.

Recent literature investigates how evoked response fields relate to neuronal disorders, such as Alzheimer’s disease, autism, dyslexia, brain tumors, and Parkinson’s disease. Furthermore, MEG has been used in conjunction with trans-axial magnetic stimulation to ameliorate abnormal brain activity (Anninos, Tsagas, Sandyk, & Derpapas, 1991).

conclusion

Magnetoencephalography is a relatively new medical imaging modality for the monitoring and imaging of human brain function. While spatial resolution is significantly lower than that of PET and fMRI, the ability to monitor neuronal activation at the millisecond time scale makes this modality, together with EEG, a unique window on the human brain. Recent developments in instrumentation have lead to the manufacture of whole-head MEG arrays with an excess of 300 magnetometers. Coupled with new data-analysis tools for mapping brain function from MEG data, these systems will lead to important new insights into the workings of the human brain with applications in both clinical and cognitive neuroscience.

KEY TERMS

ANOVA and ANCOVA: Analysis of variance or covariance is a collection of statistical models and their associated procedures that compare means by splitting the overall observed variance into different parts.

Current Dipole: Popular source model in MEG, representing a point’s current source. It is a convenient representation for the coherent activation of a large number of pyramidal cells, possibly extending over a few square centimeters of gray matter.

EEG: Electroencephalography measures neu-ronal activity by recording electrical potentials with electrodes attached on the human scalp. The resulting waveforms are used to localize brain activity and assess brain damage, epilepsy, or even in some cases brain death.

FMRI: Functional magnetic resonance imaging uses powerful magnets to create a field that resonates the nuclei of atoms in the body. The oscillating atoms emit radio signals that are converted by a computer into 3-D images of the human body and cerebral blood flow.

LCMV Beamformer: Linearly constrained minimum-variance beamformer applies spatial filtering to sensor array data to discriminate between signals from a location of interest and those originating elsewhere. In the application to MEG, the goal is to find a spatial filter that minimizes the output power of the beamformer subject to a unity gain constraint at the desired location on the brain.

MUSIC: Multiple signal classification is a localization algorithm that uses the subspace correlation between the data and model subspace to identify the origin of signals. It is often used in MEG to estimate the location, orientation, and strength of current dipoles.

PET: Positron emission tomography is a non-invasive imaging modality that measures the distribution of radioactive-labeled molecules inside a biological system. By using molecular probes that have different rates of uptake depending on the type of tissue involved, PET can localize lesions, and detect regional blood flow and gene expression among others.

Next post:

Previous post: