Multi-core beamformers: Derivation, limitations and improvements
Highlights
► Multi-core beamformer methods known as DCBF and eDCBF are explored. ► Expressions relating true sources to the DCBF/eDCBF reconstructions are derived. ► Location bias of the DCBF and eDCBF source search algorithms is analyzed. ► Rigorous requirements for the applicability of both methods are formulated. ► Performance in non-ideal practical situations is examined.
Introduction
Linear adaptive spatial filters are becoming popular source-imaging tools for electroencephalography (EEG) and magnetoencephalography (MEG) studies. These filters, also known as beamformers, reconstruct electrical activity in locations inside the brain based on the signals recorded by an array of sensors positioned outside the head (MEG) or mounted on the head surface (EEG). In particular, minimum variance beamformers (Greenblatt et al., 2005, Herdman and Cheyne, 2009, Huang et al., 2004, Robinson and Vrba, 1999, Sekihara and Nagarajan, 2008, Van Veen et al., 1997) proved to be very effective in a variety of practical situations for estimating neural generators. Mathematically the problem of reconstructing electrical sources inside the brain based on the fields observed outside the head is ill posed and cannot be solved without making additional assumptions about the sources (Baillet et al., 2001, Greenblatt et al., 2005). In particular, commonly used Linearly Constrained Minimum Variance (LCMV) filters (Van Veen et al., 1997), Spatial Aperture Magnetometry (SAM) filters (Robinson and Vrba, 1999) and their variations (Herdman and Cheyne, 2009, Huang et al., 2004, Sekihara and Nagarajan, 2008) are based on the assumption that the sources are uncorrelated or statistically orthogonal.
In practice this assumption is often violated because when a cognitive or perceptual task is performed many brain regions are involved and often exhibit correlated activity. A commonly described example of source correlation occurs as a result of bilateral sensory areas being activated simultaneously for visual or auditory stimuli (see for example Quraan and Cheyne, 2010, Herdman et al., 2003), especially in situations where synchronized activations are sustained for long time periods, as seen in the auditory steady state response (ASSR) measurements (Herdman et al., 2003).
Significant source correlations adversely affect performance of the conventional minimum variance beamformers. In particular, correlated sources tend to cancel each other leading to decreased signal-to-noise ratio (SNR) and distorted time courses (Hillebrand and Barnes, 2005, Sekihara et al., 2002, Van Veen et al., 1997). Technically, this happens because the filter weight coefficients are found independently for each location, implicitly assuming that the source at the target location is the only one. In other words, conventional beamformers are “single-source”. To improve the beamformer performance in the presence of correlated activity, other sources should somehow be accounted for in the filter-weight calculations, making the beamformer “multisource”.
One way to implement multisource beamformers is to derive the weights using a linear combination of the lead fields of possible sources (Brookes et al., 2007). This approach allowed accurate localization of a pair of highly correlated sources but it becomes computationally expensive with an exponentially increasing amount of calculations, as the number of potential sources grows. Diwakar et al. (2011a) suggested a modification of this method, which treats a pair of the source lead fields as a single higher-dimensional lead field and also reduces the number of calculations involved. The authors called this a dual-core beamformer (DCBF).
Another approach is the multiple constrained minimum variance (MCMV) beamformer, originating from the field of radar detection (Frost, 1972) and successfully applied to the EEG/MEG inverse problem in the recent years (Dalal et al., 2006, Hui et al., 2010, Moiseev et al., 2011, Popescu et al., 2008, Quraan and Cheyne, 2010). The MCMV beamformer weights are calculated under the constraint that the filter gains for signals originating from other source locations, or even from extended brain regions are (approximately) zero. This way the correlated interference at the target location, which is the main cause of problems for the single-source beamformers, is eliminated. A newer version of DCBF called the “enhanced” DCBF (eDCBF, Diwakar et al., 2011b) also imposes such constraints on the weights, and therefore belongs to the MCMV family. Although multi-source beamformers proved successful for reconstruction of bilateral activations of visual and auditory sensory areas (Popescu et al., 2008, Quraan and Cheyne, 2010) and in functional connectivity analysis (Hui et al., 2010), they have not become widely accepted practice likely due to at least two reasons.
First, in most cases the source locations are not known a priori, and the beamformer itself should be used to find them. For the single-source filters this is done by an exhaustive search over the entire brain volume. In multi-source beamforming such “brute force” approaches are currently unfeasible given the typical computing power in most laboratories, as the number of dimensions of the searched parameter space grows exponentially with the number of sources.
Second, the choice of brain activity measure, or the localizer function used to find sources is not obvious. The localizer function should reach its maximum when the true source parameters are matched. For the single-source beamformers such unbiased localizers are known, for example, the pseudo-Z ratio and the neural activity index (Greenblatt et al., 2005, Sekihara et al., 2005). In the multi-source cases various forms of the single-source localizers were still applied (Dalal et al., 2006, Popescu et al., 2008, Quraan and Cheyne, 2010) although their unbiased property was not established. Recently several unbiased multi-source localizers were derived by Moiseev et al. (2011).
The present paper investigates the validity of DCBF and eDCBF (Diwakar et al., 2011a, Diwakar et al., 2011b) methods. Each of them addresses the above problems differently and suggests its own localizer function, source search algorithm, and amplitudes and orientations reconstruction procedure. Although these suggested solutions could improve computational speed, the justification given is based on the single-source not multi-source results, and is mainly supported by numerical experiments. In this work, we investigated DCBF and eDCBF methods analytically. For the general case of arbitrary number of correlated sources, we analyzed the DCBF and eDCBF localizer functions to verify that they are unbiased, and tested the accuracy of amplitudes, orientations, and correlations. However, we did not investigate the DCBF/eDCBF source search algorithms as these are general numerical methods not specific to the bioelectromagnetic inverse problem. Furthermore, we looked at non-ideal situations where the source localization or covariance matrix estimates might be approximate. Analytical results were confirmed using computer simulations.
Section snippets
Minimum variance linear filter solutions
The following notation is used throughout this paper. Vectors and matrices are specified in lower and upper-case bold letters respectively (i.e. a and A), while scalar quantities, including the components of vectors and matrices — in regular letters (for example, t, ai, Aij, P). Additionally, we use the symbol “hat” (“^”) to distinguish an estimate of some quantity rather than its true value, which might be not known.
Let b(t) denote the M-dimensional column vector of the EEG or MEG sensor
Simulations
We evaluated the validity of DCBF and eDCBF for a pair of point sources for differing levels of SNR and inter-source correlations. All DCBF and eDCBF modeling was based on data for a 151-channel axial gradiometer MEG system (CTF/VSM). A real adult human subject's head shape, approximated with multiple local spheres, and spontaneous brain noise N collected from the same subject were used. We first considered two cases where source locations were fixed, and performed DCBF/eDCBF reconstructions
DCBF amplitudes and orientations — exact locations
The results for simulations with zero source localization errors are presented in Fig. 2. In case 1 (distant sources — see Table 1), original DCBF showed decreasing errors in source orientation with higher SNR and correlation (Fig. 2A). Note that reasonable accuracy, where the error became smaller than 10 degrees, was only achieved at fairly large SNRs (SNR > 0.5), especially for source #2. Same conclusions can be made about the amplitudes (Fig. 2B). For sources < 60 nA ∗ m and low SNR (< 0.5),
Discussion
DCBF and eDCBF (Diwakar et al., 2011a, Diwakar et al., 2011b) are minimum variance adaptive spatial filters for EEG/MEG, designed to reconstruct correlated source activity. In the current work, we investigated the limits of applying both approaches. Using exact analytical expressions, we established the necessary conditions for the DCBF and eDCBF approaches to be valid, and evaluated what happened when the conditions were violated.
First, we analyzed the source localizer functions in order to
Conclusions
In this work we investigated the accuracy and limitations of DCBF and eDCBF methods. With respect to the proposed source localizer functions, we showed that the DCBF localizer (the “pseudo-Z score”, or “K-localizer”) is unbiased. The localizer used by eDCBF (the “activity index”) is biased. Therefore, we advise against the use of the activity index, since alternative unbiased multi-source localizers exist.
With respect to the reconstructed source orientations, amplitudes and correlations, our
Acknowledgment
The Canadian Foundation for Innovation (CFI) to Urs Ribary, the Behavioral and Cognitive Neuroscience Institute (BCNI), and the Down Syndrome Research Foundation, British Columbia, Canada contributed to this research. The Michael Smith Foundation for Health Research supported this research through a Scholar Award granted to Dr. Herdman. The authors also wish to thank Ms. Nadya Moisseeva and Ms. Eleanor Stewart for their invaluable help in preparation of the manuscript.
References (25)
- et al.
Beamformer reconstruction of correlated sources using a modified source model
NeuroImage
(2007) - et al.
Optimising experimental design for MEG beamformer imaging
NeuroImage
(2008) - et al.
Investigating spatial specificity and data averaging in MEG
NeuroImage
(2010) - et al.
Dual-core beamformer for obtaining highly correlated neuronal networks in MEG
NeuroImage
(2011) - et al.
Accurate reconstruction of temporal correlation for neuronal sources using the enhanced dual-core MEG beamformer
NeuroImage
(2011) - et al.
Determination of activation areas in the human auditory cortex by means of synthetic aperture magnetometry
NeuroImage
(2003) - et al.
Beamformer analysis of MEG data
Int. Rev. Neurobiol.
(2005) - et al.
Identifying true cortical interactions in MEG using the nulling beamformer
NeuroImage
(2010) - et al.
Application of multi-source minimum variance beamformers for reconstruction of correlated neural activity
NeuroImage
(2011) - et al.
Reconstruction of correlated brain activity with adaptive spatial filters in MEG
NeuroImage
(2010)
Localization bias and spatial resolution of adaptive and non-adaptive spatial filters for MEG source reconstruction
NeuroImage
Electromagnetic brain mapping
IEEE Signal Process. Mag.
Cited by (21)
Localization of brain signals by alternating projection
2024, Biomedical Signal Processing and ControlSubspace based Multiple Constrained Minimum Variance (SMCMV) beamformers
2022, Biomedical Signal Processing and ControlLocalization of brain activity from EEG/MEG using MV-PURE framework
2021, Biomedical Signal Processing and ControlCitation Excerpt :In Section 2 we introduce the notation, the EEG/MEG forward model, and the definitions of an unbiased activity index and its spatial resolution. We also provide definitions of full-rank activity indices introduced in [10] along with their generalizations proposed in [30]. In Section 3 we propose novel reduced-rank activity indices, prove that they are unbiased, and discuss their spatial resolution as a function of their rank.
Multiple constrained minimum variance beamformer (MCMV) performance in connectivity analyses
2020, NeuroImageCitation Excerpt :We will be using terms “reconstruction” and “estimation” as synonyms further on, although strictly speaking for real data the ground truth is never known and “estimation” is the only option. One popular inverse model method is linearly constrained minimum variance (LCMV) beamforming (Van Veen et al., 1997; Vrba and Robinson, 2001; Sekihara et al., 2002, 2007; Robinson, 2004; Herdman and Douglas, 2009; Moiseev and Herdman, 2013) which has been shown to be very reliable in localizing sources and estimating brain activity (Sekihara et al., 2002; Dalal et al., 2008; Murzin et al., 2011; Jonmohamadi et al., 2014). Inverse modeling, however, is an ill posed problem (Michel et al., 2004; Greenblatt et al., 2005).
Minimum variance beamformer weights revisited
2015, NeuroImageCitation Excerpt :si(θ0i, t) and ν(t) are assumed to be uncorrelated zero-mean stationary random processes: 〈si(θ0i, t)〉 = 0, 〈ν(t)〉 = 0, 〈si(θ0i, t)ν〉 = 0, where the angle brackets denote statistical averaging. Eq. (3) or its sub-cases is commonly used by all popular minimum variance filters, including single-source scalar and vector beamformers (Robinson and Vrba, 1999; Van Veen et al., 1997; Sekihara and Nagarajan, 2008; Huang et al., 2004), evoked beamformers (Robinson, 2004; Cheyne et al., 2007), as well as multi-source versions of those (Brookes et al., 2007; Dalal et al., 2006; Diwakar et al., 2011; Moiseev et al., 2011, Moiseev and Herdman, 2013). Sometimes, additional normalizations are applied to the weights themselves, or to the lead fields in Eq. (3), or to both (see Sekihara and Nagarajan, 2008 for examples).