Here, we demonstrated a computationally efficient method of a wavelet-based, anatomically constrained MNE to create an image of signal frequency power and phase on the cortical surface. In particular, we used individual anatomy (head and brain) to further guide the source estimation, and display it on the cortex. The technique visualizes short-lasting activity including both the spectral power over some frequency band and the phase relationship, either between brain regions or between a brain region and the stimulus.
We emphasize the importance of choosing and calculating the appropriate noise covariance matrix. For spontaneous activities, such as 10-Hz alpha rhythm in our results, the noise refers to the noise observed extraneous from the subject’s brain activity. Therefore, empty room recordings were used to calculate this noise covariance matrix.
For the noise normalization, the term “noise covariance” actually refers to the covariance of the prestimulus interval, that is, the baseline covariance. An incorrect choice of the baseline covariance causes either overestimation or underestimation of the baseline activity as well as the possibility of mislocalization of the statistical significance calculated by noise normalization. In addition, the baseline covariance should always be computed from the raw data to increase the number of samples available to the estimator. A baseline covariance matrix estimated from averaged data always undercharacterizes the variability of the measurement and thus leads to overestimated significance in the statistical tests. This is true because the covariance of the averaged data does not account for trial-to-trial variations. Therefore, raw data in appropriate “baseline” conditions have to be chosen appropriately for noise normalization.
Comparing our simulations and the results from real data, we validated that the reported technique can be used to detect frequency-specific responses at signal to noise ratio (SNR) as small as 0.01, if up to 100 trials of measurements are available and the activity is phase-locked to the stimulus (evoked). However, if the activity is not phase locked to the stimulus (induced), the SNR must be greater than 1, a limitation that is true for measuring induced spectral power. Nevertheless, the magnitude of the F statistic and PLV differed in the simulations and the actual MEG measurements. Potentially, this may be due to (1) anatomical variations in the neural sources (not precisely perpendicular orientations of dipole sources to the local cortical surface), (2) anatomical information errors from MRI, and (3) time-varying activities in real brain responses (changes of spatial distribution and temporal duration of the oscillations). In the Bayesian estimation theory framework, this can be further improved by providing more accurate prior information regarding both the spatial and temporal properties of the brain activations, such as more accurate quantitative description between BOLD fMRI and MEG measurements. This is a current research topic our laboratory is actively investigating. It is hoped that this method will make such research possible.
Comparison with other source estimation methods
Currently, several methods have been proposed to map cortical oscillations, each of which has a particular strength for its application, but none of which meets all the requirements proposed here. Minimum current estimate (MCE), for example (Matsuura and Okabe, 1995
; Uutela et al., 1999
), assumes sources with exponential probability distribution function. Specifically, Jensen and Vanni (2002)
proposed to use MCE to measure simultaneous cortical oscillations, based on the minimum L1
-norm estimate, using complex Fourier components from sections of raw MEG signals. Although it is limited to the analysis of spontaneous cortical oscillations, the method could be extended to include measurements of synchronization based on the phase of the Fourier components. It is, however, limited in that it does not have a high-temporal resolution to measure fast changing oscillations and requires a high SNR. Source localizations using beamformer techniques (Van Veen et al., 1997
; Vrba et al., 1999
) have been also applied to study neural synchrony, such as dynamic imaging of coherent sources (DICS) (Gross et al., 2001
). This innovative technique is a promising method of detecting synchronizations between neural networks in the brain, based on source estimation using a spatial filter of coherence. Like the MCE technique, it is limited, however, in that it requires a relatively high SNR and makes the assumption of stationary data, and consequently does not measure short lasting synchronizations. Hashimoto et al. (2001)
, Sekihara et al. (2001)
, and Sekihara et al. (2002)
recently demonstrated that the stationary data assumption is not necessarily strict.
The use of the MNE to image cortical oscillations has also been proposed by others (David et al., 2002
; Hauk et al., 2002
). Hauk et al. (2002)
compared the MNE to a current source density method using EEG data sets with a wavelet-based frequency domain analysis. Others (David et al., 2002
; Lachaux et al., 2002
) have recently performed simulations on various versions of the MNE. They concluded that MNE is a preferred method compared to these other minimum L2
-norm inverse approaches. Our method differs in several ways. First, our solution computes single trial epochs, or uses clips of spontaneous activity, to create an averaged power spatiotemporal map. Second, we use a BEM to use the features of the actual head of the individual subject’s MRI to create a power sources estimate directly on the cortical surface, utilizing a cortical constraint. Third, we also include the ability to create a phase synchronization map of the brain on the cortical surface. In this case, we show how the phase relation of the stimulus to the response can be created, but the same principle can be used to create maps showing synchronization between brain regions, provided the reference temporal course from one specific brain region. And last, the utilization of prior neural activation information from fMRI can help for correct power localization. This is especially important during the localization of the evoked auditory responses because of highly convoluted cortical surface at the temporal lobe. The anatomy challenges the traditional MNE to give focal and precise source estimations. Without fMRI priors, it is easy to lead to superficial source estimates away from Heschl’s gyrus.
Our method (sdSPM) has similarities to that proposed by David et al. (2002)
in that both source estimation methods use a wavelet-based spectral analysis and a MNE. Our method differs in that we calculate the wavelet-based spectral power based on the power in source space (on the brain surface) across trains, instead of calculating the single trial power and single trial phase locking maps as used by David et al. We have also proposed a method of producing a spectral dSPM with noise normalization of the MNE.
Several limitations of this technique are illustrated in this report. The algorithm’s localizing ability is often limited, especially with activity in and around the Sylvian fissure, a common problem with imaging peri-Sylvian activity generally and auditory activity specifically. We can see in our results that probable ‘phantom’ sources are seen in the posterior inferior frontal lobe, insula, and the middle temporal gyrus. These sources, even with actual measurements, are likely phantom as they are seen in the simulations as well due to the close anatomical proximity to Heschl’s gyrus. Due to the anatomical convolution of cortical surface, traditional MNE without a priori activation loci may lead to mislocalization of auditory neural activities. Using fMRI, which provides a higher spatial resolution brain activation estimate, it is possible to constrain the localization process to avoid overestimation of the power of neuromagnetic sources in traditional MNE. For phase-locking value and noise-normalized power estimates, it appears that a priori information could be less essential. This is because of the intrinsic normalization of PLV calculation by disregarding the amplitude of dipole estimates in an individual trial. The noise normalization also compensates for the biases on the magnitude of the estimates. Therefore, PLV and dSPM localization shows similar sensitivity for both superficial sources on the gyri and deep sources in the sulci. Another limitation of our approach is that no deep gray nuclei activity is included in the source model. Thalamocortical interaction is probably responsible for the generation and maintenance of oscillations in the neocortex. Our group is actively investigating this issue and it will be a focus for an upcoming study.
Although both PLV and noise-normalized power can correct biases toward superficial cortical activities, potential overcorrection may occur, as we found in both the simulations and real data that the insula was probably falsely activated. For simultaneous high sensitivity and specificity, fMRI-weighted MNE can be a candidate tool for auditory response localization.
The new method detailed here is a general framework and can be extended to include other imaging technologies including PET, optical imaging, and transcranial Doppler. Any functional information that can be localized can be included in the R matrix and used as prior information and thus improve the localizing ability of the MEG or EEG data.