Deep Source Localization with Magnetoencephalography Based on Sensor Array Decomposition and Beamforming

Abstract

In recent years, the source localization technique of magnetoencephalography (MEG) has played a prominent role in cognitive neuroscience and in the diagnosis and treatment of neurological and psychological disorders. However, locating deep brain activities such as in the mesial temporal structures, especially in preoperative evaluation of epilepsy patients, may be more challenging. In this work we have proposed a modified beamforming approach for finding deep sources. First, an iterative spatiotemporal signal decomposition was employed for reconstructing the sensor arrays, which could characterize the intrinsic discriminant features for interpreting sensor signals. Next, a sensor covariance matrix was estimated under the new reconstructed space. Then, a well-known vector beamforming approach, which was a linearly constraint minimum variance (LCMV) approach, was applied to compute the solution for the inverse problem. It can be shown that the proposed source localization approach can give better localization accuracy than two other commonly-used beamforming methods (LCMV, MUSIC) in simulated MEG measurements generated with deep sources. Further, we applied the proposed approach to real MEG data recorded from ten patients with medically-refractory mesial temporal lobe epilepsy (mTLE) for finding epileptogenic zone(s), and there was a good agreement between those findings by the proposed approach and the clinical comprehensive results.

Partial Least Square Aided Beamforming Algorithm in Magnetoencephalography Source Imaging

“…In this section, we first verify the source imaging algorithm through the experiments based on simulated data, generated by a sinc function plus Gaussian noise. The proposed method is compared with three well-known beamforming methods: linearly constrained minimum variance (LCMV) (Van Veen et al, 1997 ), dynamic imaging of coherent sources (DICS) (Groß et al, 2001 ), and modified LCMV with iterative matrix decomposition (mLCMV) (Hu et al, 2017 ). Because the new method combines PLS with beamforming, pLCMV is regarded as an acronym of the new method.…”

“…With the selection of the input variables X and the response variables Y , the PLS analysis generates multiple variants in different application scenarios. Recent studies (Sekihara et al, 2002b ; Brookes et al, 2007 ; Hu et al, 2017 ) ignore the correlations between different brain regions when reconstructing the input matrix X in the MEG source imaging. Since PLS is a supervised learning, the first step in using this method is to divide all samples into multiple classes.…”

“…The main reason for this problem is that the sensor array geometry is used directly to estimate the covariance matrix. Recently, an iterative spatiotemporal signal decomposition method has been used to modify the vector beamforming technique, and has been successfully applied to source localization for MEG signals (Hu et al, 2017 ). Although the approach has achieved high spatial accuracy, the correlations of signals from different brain regions are ignored when the components of the MEG arrays are extracted.…”

“…We then reconstructed the sensor arrays based on the components and loadings, and used the vector beamforming technique to estimate the source model. The newly proposed source imaging approach for MEG recordings was first validated on simulated data, and compared with three other well-known beamforming methods, linearly constrained minimum variance (LCMV) (Van Veen et al, 1997 ), dynamic imaging of coherent sources (DICS) (Groß et al, 2001 ), and modified LCMV with iterative matrix decomposition (mLCMV) (Hu et al, 2017 ). Since these methods belong to the beamforming family, the basic assumption in this study is the same as assumption underlying the minimum variance beamforming.…”

Fast oscillations >40Hz localize the epileptogenic zone: an electrical source imaging study using high-density electroencephalography

“…Most previous work in source localization of FOs in MEG used the Beamformer technique, as it is assumed to be able to detect distributed and deep sources [Hu et al, 2017]. Yet Beamformer is not a source localization method in its proper sense, but corresponds rather to a statistical dipole scanning approach, iteratively assessing how likely it would be to fit an equivalent current dipole at a specific position in a 3D grid covering the brain [Hillebrand et al, 2005].…”