Statistical performance analysis of signal variance-based dipole models for MEG/EEG source localization and detection

A set of dipole fitting algorithms that incorporate different assumptions about the variability of the signal component into their mathematical models is presented and analyzed. Dipole fitting is performed by minimizing the squared error between the selected data model and available data. Dipole models based on moments that have 1) constant amplitude and orientation, 2) variable amplitude and fixed known orientation, 3) variable amplitude and fixed unknown orientation, and 4) variable amplitude and variable orientation are considered. The presence of a dipolar source is determined by comparing the fractional energy explained by the dipole model to a threshold. Source localization is accomplished by searching to find the location that explains the largest fractional signal energy using a dipole model. Expressions for the probability of a false positive decision and probability of correct detection are derived and used to evaluate the effect of variability in the dipole on performance and to address the effects of model mismatch and location errors. Simulated and measured data experiments are presented to illustrate the performance of both detection and localization methods. The results indicate that models which account for variance outperform the constant orientation and magnitude model even when the number of observations is relatively small and the signal of interest contains a very modest variance component.     

Maximum-likelihood estimation of low-rank signals for multiepoch MEG/EEG analysis

A maximum-likelihood-based algorithm is presented for reducing the effects of spatially colored noise in evoked response magneto- and electro-encephalography data. The repeated component of the data, or signal of interest, is modeled as the mean, while the noise is modeled as the Kronecker product of a spatial and a temporal covariance matrix. The temporal covariance matrix is assumed known or estimated prior to the application of the algorithm. The spatial covariance structure is estimated as part of the maximum-likelihood procedure. The mean matrix representing the signal of interest is assumed to be low-rank due to the temporal and spatial structure of the data. The maximum-likelihood estimates of the components of the low-rank signal structure are derived in order to estimate the signal component. The relationship between this approach and principal component analysis (PCA) is explored. In contrast to prestimulus-based whitening followed by PCA, the maximum-likelihood approach does not require signal-free data for noise whitening. Consequently, the maximum-likelihood approach is much more effective with nonstationary noise and produces better quality whitening for a given data record length. The efficacy of this approach is demonstrated using simulated and real MEG data.     

Estimating brain conductivities and dipole source signals with EEG arrays

Techniques based on electroencephalography (EEG) measure the electric potentials on the scalp and process them to infer the location, distribution, and intensity of underlying neural activity. Accuracy in estimating these parameters is highly sensitive to uncertainty in the conductivities of the head tissues. Furthermore, dissimilarities among individuals are ignored when standarized values are used. In this paper, we apply the maximum-likelihood and maximum a posteriori (MAP) techniques to simultaneously estimate the layer conductivity ratios and source signal using EEG data. We use the classical 4-sphere model to approximate the head geometry, and assume a known dipole source position. The accuracy of our estimates is evaluated by comparing their standard deviations with the Cramér-Rao bound (CRB). The applicability of these techniques is illustrated with numerical examples on simulated EEG data. Our results show that the estimates have low bias and attain the CRB for sufficiently large number of experiments. We also present numerical examples evaluating the sensitivity to imprecise assumptions on the source position and skull thickness. Finally, we propose extensions to the case of unknown source position and present examples for real data.     

A multiresolution framework to MEG/EEG source imaging

A new method based on a multiresolution approach for solving the ill-posed problem of brain electrical activity reconstruction from electroencephaloram (EEG)/magnetoencephalogram (MEG) signals is proposed in a distributed source model. At each step of the algorithm, a regularized solution to the inverse problem is used to constrain the source space on the cortical surface to be scanned at higher spatial resolution. We present the iterative procedure together with an extension of the ST-maximum a posteriori method [1] that integrates spatial and temporal a priori information in an estimator of the brain electrical activity. Results from EEG in a phantom head experiment with a real human skull and from real MEG data on a healthy human subject are presented. The performances of the multiresolution method combined with a nonquadratic estimator are compared with commonly used dipolar methods, and to minimum-norm method with and without multiresolution. In all cases, the proposed approach proved to be more efficient both in terms of computational load and result quality, for the identification of sparse focal patterns of cortical current density, than the fixed scale imaging approach.     

Spatiotemporal EEG/MEG source analysis based on a parametric noise covariance model

A method is described to incorporate the spatiotemporal noise covariance matrix into a spatiotemporal source analysis. The essential feature is that the estimation problem is split into two parts. First, a model is fitted to the observed noise covariance matrix. This model is a Kronecker product of a spatial and a temporal matrix. The spatial matrix models the spatial covariances by a function dependent on sensor distance. The temporal matrix models the temporal covariances as lag dependent. In the second part, sources are estimated given this noise model, which can be done very efficiently due to the Kronecker formulation. An application to real electroencephalogram (EEG) data shows that the noise model fits the data very well. Simulation results show that the resulting source estimates are more precise than those obtained from a standard analysis neglecting the noise covariance. In addition, the estimated standard errors of the source parameter estimates are far more precise than those obtained from a standard analysis. Finally, the source parameter standard errors are used to investigate the effects of temporal sampling. It is shown that increasing the sampling by a factor x, decreases the standard errors of all source parameters with the square root of x.     

A resampling method for estimating the signal subspace of spatio-temporal EEG/MEG data

Source localization using spatio-temporal electroencephalography (EEG) and magnetoencephalography (MEG) data is usually performed by means of signal subspace methods. The first step of these methods is the estimation of a set of vectors that spans a subspace containing as well as possible the signal of interest. This estimation is usually performed by means of a singular value decomposition (SVD) of the data matrix: The rank of the signal subspace (denoted by r) is estimated from a plot in which the singular values are plotted against their rank order, and the signal subspace itself is estimated by the first r singular vectors. The main problem with this method is that it is strongly affected by spatial covariance in the noise. Therefore, two methods are proposed that are much less affected by this spatial covariance, and old and a new method. The old method involves prewhitening of the data matrix, making use of an estimate of the spatial noise covariance matrix. The new method is based on the matrix product of two average data matrices, resulting from a random partition of a set of stochastically independent replications of the spatio-temporal data matrix. The estimated signal subspace is obtained by first filtering out the asymmetric and negative definite components of this matrix product and then retaining the eigenvectors that correspond to the r largest eigenvalues of this filtered data matrix. The main advantages of the partition-based eigen decomposition over prewhited SVD is that 1) it does not require an estimate of the spatial noise covariance matrix and 2b) that it allows one to make use of a resampling distribution (the so-called partitioning distribution) as a natural quantification of the uncertainty in the estimated rank. The performance of three methods (SVD with and without prewhitening, and the partition-based method) is compared in a simulation study. From this study, it could be concluded that prewhited SVD and the partition-based eigen decomposition perform equally well when the amplitude time series are constant, but that the partition-based method performs better when the amplitude time series are variable.     

Spatio-temporal reconstruction of bilateral auditory steady-state responses using MEG beamformers

A rapidly growing number of neuromagnetic studies focus on the analysis of auditory steady-state responses (ASSR) in relation to a diverse array of factors including age, selective attention, and presence of psychopathology. The objectives of these studies require accurate spatio-temporal estimation of the underlying neural generators, a challenging task due to the relatively low signal strength and high correlation between bilateral auditory cortical sources. This paper evaluates the performance of two beamforming schemes that can potentially overcome such difficulties: 1) the linearly constrained minimum variance beamformer with partial sensor coverage (LCMV-PSC), and 2) the multiple constrained minimum-variance beamformer with coherent source region suppression (MCMV-CSRS). Simulation experiments are conducted to assess the impact of source parameters on the reconstruction accuracy. The results indicate that the LCMV-PSC method is prone to localization errors that essentially occur along medio-lateral directions, increase with source depth, and are associated to amplitude and phase distortions of the estimated time courses of activity. Comparatively, the MCMV-CSRS method exhibits precise localization and minimal amplitude and phase distortion for a broad range of relative interferer's positions within the suppression region. The results from the numerical experiments are validated on real magnetoencephalographic (MEG) data collected from a 40-Hz ASSR paradigm.     

Estimating evoked dipole responses in unknown spatially correlatednoise with EEG/MEG arrays

We present maximum likelihood (ML) methods for estimating evoked dipole responses using electroencephalography (EEG) and magnetoencephalography (MEG) arrays, which allow for spatially correlated noise between sensors with unknown covariance. The electric source is modeled as a collection of current dipoles at fixed locations and the head as a spherical conductor. We permit the dipoles' moments to vary with time by modeling them as linear combinations of parametric or nonparametric basis functions. We estimate the dipoles' locations and moments and derive the Cramer-Rao bound for the unknown parameters. We also propose an ML based method for scanning the brain response data, which can be used to initialize the multidimensional search required to obtain the true dipole location estimates. Numerical simulations demonstrate the performance of the proposed methods     

Convergence analysis of finite alphabet beamformers for digital cochannel signals

In this paper, we examine the convergence behavior of finite alphabet (FA) beamformers. The two most popular implementations of FA beamformers for digital communication signals are the iterative least-squares with projection (ILSP) and the minimum mean-square error (MMSE) beamformer. To facilitate the analysis, for the binary communications case, we derive closed-form expressions for the mean weight vector, signal-to-noise ratio, and signal-to-interference ratio for both the ILSP and MMSE beamformers in terms of the bit-error rate (BER) performance at each iteration. Next, we generalize the analysis for the M-ary pulse-amplitude modulation and M-ary phase-shift keying cases. We show that both ILSP and MMSE beamformers have previously unreported bias terms in the array response vectors which are functions of the BER for each iteration. Furthermore, as the BER becomes arbitrarily small, we show that our solutions converge to the well-known asymptotic expressions widely published in the literature. Next, we provide a geometric interpretation of the effects of the noise bias vector in terms of angles between subspaces. Based on our analysis, we were able to develop necessary and sufficient conditions for convergence in the mean. We conclude with Monte-Carlo simulations to validate our analysis.     

Recursive MUSIC: A framework for EEG and MEG source localization

The multiple signal classification (MUSIC) algorithm can be used to locate multiple asynchronous dipolar sources from electroencephalography (EEG) and magnetocncephalography (MEG) data. The algorithm scans a single-dipole model through a three-dimensional (3-D) head volume and computes projections onto an estimated signal subspace. To locate the sources, the user must search the head volume for multiple local peaks in the projection metric. This task is time consuming and subjective. Here, the authors describe an extension of this approach which they refer to as recursive MUSIC (R-MUSIC). This new procedure automatically extracts the locations of the sources through a recursive use of subspace projections. The new method is also able to locate synchronous sources through the use of a spatio-temporal independent topographies (IT) model. This model defines a source as one or more nonrotating dipoles with a single time course. Within this framework, the authors are able to locate fixed, rotating, and synchronous dipoles. The recursive subspace projection procedure that they introduce here uses the metric of canonical or subspace correlations as a multidimensional form of correlation analysis between the model subspace and the data subspace, by recursively computing subspace correlations, the authors build up a model for the sources which account for a given set of data. They demonstrate here how R-MUSIC can easily extract multiple asynchronous dipolar sources that are difficult to find using the original MUSIC scan. The authors then demonstrate R-MUSIC applied to the more general IT model and show results for combinations of fixed, rotating, and synchronous dipoles     

EEG/MEC error bounds for a static dipole source with a realistichead model

We derive Cramer-Rao bounds (CRBs) on the errors of estimating the parameters (location and moment) of a static current dipole source using data from electro-encephalography (EEG), magneto-encephalography (MEG), or the combined EEG/MEG modality. We use a realistic head model based on knowledge of surfaces separating tissues of different conductivities obtained from magnetic resonance (MR) or computer tomography (CT) imaging systems. The electric potentials and magnetic field components at the respective sensors are functions of the source parameters through integral equations. These potentials and field are formulated for solving them by the boundary or the finite element method (BEM or FEM) with a weighted residuals technique. We present a unified framework for the measurements computed by these methods that enables the derivation of the bounds. The resulting bounds may be used, for instance, to choose the best configuration of the sensors for a given patient and region of expected source location. Numerical results are used to demonstrate an application for showing expected accuracies in estimating the source parameters as a function of its position in the brain, based on real EEG/MEG system and MR or CT images     

Combined MEG and EEG source imaging by minimization of mutual information

Though very frequently assumed, the necessity to operate a joint processing of simultaneous magnetoencephalography (MEG) and electroencephalography (EEG) recordings for functional brain imaging has never been clearly demonstrated. However, the very last generation of MEG instruments allows the simultaneous recording of brain magnetic fields and electrical potentials on the scalp. But the general fear regarding the fusion between MEG and EEG data is that the drawbacks from one modality will systematically spoil the performances of the other one without any consequent improvement. This is the case for instance for the estimation of deeper or radial sources with MEG. In this paper, we propose a method for a cooperative processing of MEG and EEG in a distributed source model. First, the evaluation of the respective performances of each modality for the estimation of every dipole in the source pattern is made using a conditional entropy criterion. Then, the algorithm operates a preprocessing of the MEG and EEG gain matrices which minimizes the mutual information between these two transfer functions, by a selective weighting of the MEG and EEG lead fields. This new combined EEG/MEG modality brings major improvements to the localization of active sources, together with reduced sensitivity to perturbations on data.     

Hierarchical clustering and filtering in half-inverse space for MEG and/or EEG hypothesis-free analysis

We propose hierarchical clustering and filtering methods for the analysis of spatio-temporal multidimensional time series, where both methods are based on a new pseudo distance. The pseudo distance is determined between orthogonal matrices, which are derived by eigenvalue decomposition of the variance-covariance matrix of the time series. Because the grouping algorithm is also important in clustering, a modified Ward method grouping criterion is used here. The filtering derives temporal similarity information between two time series, providing information that cannot be evaluated by the clustering. If the time series to be clustered and filtered cannot be obtained directly, different time series reflecting the original time series are used instead. There exists a transform between the time series, and hence, scaling distortion occurs. We also propose a scaling normalization method. As an application example, we present an analysis of a multichannel magnetoencephalography (MEG) and/or electroencephalography (EEG) time series. Each of the MEG and EEG generations is a transform from the same electrophysiological brain activity. We applied these methods to sound localization-related MEG time series and evaluated their effectiveness. These methods may be useful for discovering similarity among many multidimensional time series without a priori information and/or hypotheses.     

Performance analysis of narrowband beamforming using fully and partial adaptive beamformers with a spherical array

This research proposes an improved narrowband partial adaptive beamformer analysis using a proposed spherical array. Comparison between the fully and the partial adaptive beamformers is given. The study is performed by investigating performance parameters like the beamformer output signal-to-noise ratio and the beamformer output signal-to-interference-plus-noise ratio both in the steady state and along adaptation. Furthermore, computational complexity and convergence speed of the proposed sensor arrangement are also analyzed and examples are given. The results demonstrate that this beamformer considerably reduces the number of complex operations and features faster convergence speed.     

Multi-way space-time-wave-vector analysis for EEG source separation

For the source analysis of electroencephalographic (EEG) data, both equivalent dipole models and more realistic distributed source models are employed. Several authors have shown that the canonical polyadic decomposition (also called ParaFac) of space-time-frequency (STF) data can be used to fit equivalent dipoles to the electric potential data. In this paper we propose a new multi-way approach based on space-time-wave-vector (STWV) data obtained by a 3D local Fourier transform over space accomplished on the measured data. This method can be seen as a preprocessing step that separates the sources, reduces noise as well as interference and extracts the source time signals. The results can further be used to localize either equivalent dipoles or distributed sources increasing the performance of conventional source localization techniques like, for example, LORETA. Moreover, we propose a new, iterative source localization algorithm, called Binary Coefficient Matching Pursuit (BCMP), which is based on a realistic distributed source model. Computer simulations are used to examine the performance of the STWV analysis in comparison to the STF technique for equivalent dipole fitting and to evaluate the efficiency of the STWV approach in combination with LORETA and BCMP, which leads to better results in case of the considered distributed source scenarios.     

An efficient algorithm for computing multishell spherical volumeconductor models in EEG dipole source localization

Computationally localizing electrical current sources of the electroencephalographic signal requires a volume conductor model which relates theoretical scalp potentials to the dipolar source located within the modeled brain. The commonly used multishell spherical model provides this source-potential relationship using a sum of infinite series whose computation is difficult. This paper provides a closed-form approximation to this sum based on an optimal fitting to the weights of the Legendre polynomials. The second-order (third-order) approximation algorithm, implemented by a provided C-routine, requires only 100 (140) floating point operations to compute a single scalp potential in response to an arbitrary current dipole located within a four-shell spherical volume conductor model. This cost of computation represents only 6.3% (8.9%) of that required by the direct method. The relative mean square error, measured by using 20,000 random dipoles distributed within the modeled brain, is only 0.29% (0.066%)     

Spatiotemporal forward solution of the EEG and MEG using network modeling

Dynamic systems have proven to be well suited to describe a broad spectrum of human coordination behavior such synchronization with auditory stimuli. Simultaneous measurements of the spatiotemporal dynamics of electroencephalographic (EEG) and magnetoencephalographic (MEG) data reveals that the dynamics of the brain signals is highly ordered and also accessible by dynamic systems theory. However, models of EEG and MEG dynamics have typically been formulated only in terms of phenomenological modeling such as fixed-current dipoles or spatial EEG and MEG patterns. In this paper, it is our goal to connect three levels of organization, that is the level of coordination behavior, the level of patterns observed in the EEG and MEG and the level of neuronal network dynamics. To do so, we develop a methodological framework, which defines the spatiotemporal dynamics of neural ensembles, the neural field, on a sphere in three dimensions. Using magnetic resonance imaging we map the neural field dynamics from the sphere onto the folded cortical surface of a hemisphere. The neural field represents the current flow perpendicular to the cortex and, thus, allows for the calculation of the electric potentials on the surface of the skull and the magnetic fields outside the skull to be measured by EEG and MEG, respectively. For demonstration of the dynamics, we present the propagation of activation at a single cortical site resulting from a transient input. Finally, a mapping between finger movement profile and EEG/MEG patterns is obtained using Volterra integrals.     

Independent component approach to the analysis of EEG and MEG recordings

Multichannel recordings of the electromagnetic fields emerging from neural currents in the brain generate large amounts of data. Suitable feature extraction methods are, therefore, useful to facilitate the representation and interpretation of the data. Recently developed independent component analysis (ICA) has been shown to be an efficient tool for artifact identification and extraction from electroencephalographic (EEG) and magnetoencephalographic (MEG) recordings. In addition, ICA has been applied to the analysis of brain signals evoked by sensory stimuli. This paper reviews our recent results in this field.     

A Spatiotemporal Framework for Estimating Trial-to-Trial Amplitude Variation in Event-Related MEG/EEG

A spatiotemporal framework for estimating trial-to-trial variability in evoked response (ER) data is presented. Spatial and temporal bases capture the aspects of the response that are consistent across trials, while the basis expansion coefficients represent the variable components of the response. We focus on the simplest case of constant spatiotemporal response shape and varying amplitude across trials. Two different constraints on the amplitude evolution are employed to effectively integrate the individual responses and improve robustness at low SNR. The linear dynamical system response constraint estimates the current trial amplitude as an unknown constant scaling of the estimate in the previous trial plus zero-mean Gaussian noise with unknown variance. The independent response constraint estimates response amplitudes across trials as independent Gaussian random variables having unknown mean and variance. We develop a generalized expectation-maximization algorithm to obtain the maximum-likelihood (ML) estimates of the signal waveform, noise covariance matrix, and unknown constraint parameters. ML source localization is achieved by scanning the likelihood over different sets of spatial bases. We demonstrate the variability estimation and source localization effectiveness of the proposed algorithms using both real and simulated ER data.