5 EEG Source Imaging and Multimodal Neuroimaging
109
uncertainty in the calculation of the cortical current sources. The fMRI activation
map, in this case, serves only to restrict the potential activity of sources in the
cortical layer by penalizing sources in fMRI-inactive regions. This is emblematic
of asymmetrical methods; one imaging modality provides the directly observed
results while the remaining method acts as a guiding influence. There are also a
few considerations which must be made for this fMRI-MNE method. First, strong
activity in the cortical reconstruction can overcome the penalty imposed by the
fMRI map. This means that the erroneous activity of sufficient strength may still
be present in the results. Second, the MRI constraint does not interact with other
noise-normalized estimates (dSPM, sLORETA, etc.). While MNE itself aims to
provide a reconstruction of cortical currents, these noise-normalization algorithms
seek to identify where cortical activity is different from baseline noise. This function
is essentially the same as fMRI, which identifies the 3D voxels that are significantly
different from background noise. Thirdly, in the above implementation, fMRI spatial
information acts as a “hard” constraints; it is assumed to be the absolute truth, in spite
of cases where fMRI “extra” sources (sources deemed active in fMRI but not EEG,
[49, 51]) can be found. Finally, there is a built-in assumption that the constraints
provided by fMRI are applicable at each time instance of the EEG measurement.
The fMRI activation map employed in this method is static, regardless of the time
point use in analysis—this means that any erroneous activity at an MRI-active voxel
will be amplified similarly to true activation. To alleviate the issue of temporal
mismatch between fMRI and EEG, Liu and He [52] proposed an EEG inversion
approach that utilized the fMRI information as time-variant spatial constraints. The
fMRI-derived prior spatial weights are adaptively varied at each time instance of
the EEG time-course based on all the EEG single-trials before averaging. On the
other hand, Daunizeau et al. addressed the fMRI “hard” constraint issue by dividing
the fMRI activation map into multiple submaps and estimating the optimal subset
of fMRI weights using model comparison approach in the parametric empirical
Bayesian framework [20]. fMRI spatial information can be utilized in a spatiotemporal specific fashion [60], employing an appropriate subset of the fMRI activation
map as spatial priors for a sliding-window of EEG time segments, thus improving
the spatial and temporal accuracy in complex and dynamic brain activity [61].
EEG-Informed fMRI
The above method has shown a common combination, wherein fMRI is used as a basis
to constrain EEG-based source localization. The lofty temporal resolution of this
approach makes it ideal for investigative or computational approaches. Oftentimes,
however, clinical and scientific researchers may instead wish to highlight or localize
a specific feature or spike in EEG activity. In these cases, they may wish to rely
more heavily upon the high spatial resolution of fMRI than the temporal speed of
EEG. This can be particularly valuable when the events targeted for characterization
fall into one of the following categories: (1) events that are largely uncontrolled or
unpredictable (investigations of epileptic activity); (2) events with a large degree of
variability; and (3) events that are invisible to fMRI alone. Serving as an converse
method to fMRI-informed EEG, results obtained by EEG-informed fMRI are similar
109
uncertainty in the calculation of the cortical current sources. The fMRI activation
map, in this case, serves only to restrict the potential activity of sources in the
cortical layer by penalizing sources in fMRI-inactive regions. This is emblematic
of asymmetrical methods; one imaging modality provides the directly observed
results while the remaining method acts as a guiding influence. There are also a
few considerations which must be made for this fMRI-MNE method. First, strong
activity in the cortical reconstruction can overcome the penalty imposed by the
fMRI map. This means that the erroneous activity of sufficient strength may still
be present in the results. Second, the MRI constraint does not interact with other
noise-normalized estimates (dSPM, sLORETA, etc.). While MNE itself aims to
provide a reconstruction of cortical currents, these noise-normalization algorithms
seek to identify where cortical activity is different from baseline noise. This function
is essentially the same as fMRI, which identifies the 3D voxels that are significantly
different from background noise. Thirdly, in the above implementation, fMRI spatial
information acts as a “hard” constraints; it is assumed to be the absolute truth, in spite
of cases where fMRI “extra” sources (sources deemed active in fMRI but not EEG,
[49, 51]) can be found. Finally, there is a built-in assumption that the constraints
provided by fMRI are applicable at each time instance of the EEG measurement.
The fMRI activation map employed in this method is static, regardless of the time
point use in analysis—this means that any erroneous activity at an MRI-active voxel
will be amplified similarly to true activation. To alleviate the issue of temporal
mismatch between fMRI and EEG, Liu and He [52] proposed an EEG inversion
approach that utilized the fMRI information as time-variant spatial constraints. The
fMRI-derived prior spatial weights are adaptively varied at each time instance of
the EEG time-course based on all the EEG single-trials before averaging. On the
other hand, Daunizeau et al. addressed the fMRI “hard” constraint issue by dividing
the fMRI activation map into multiple submaps and estimating the optimal subset
of fMRI weights using model comparison approach in the parametric empirical
Bayesian framework [20]. fMRI spatial information can be utilized in a spatiotemporal specific fashion [60], employing an appropriate subset of the fMRI activation
map as spatial priors for a sliding-window of EEG time segments, thus improving
the spatial and temporal accuracy in complex and dynamic brain activity [61].
EEG-Informed fMRI
The above method has shown a common combination, wherein fMRI is used as a basis
to constrain EEG-based source localization. The lofty temporal resolution of this
approach makes it ideal for investigative or computational approaches. Oftentimes,
however, clinical and scientific researchers may instead wish to highlight or localize
a specific feature or spike in EEG activity. In these cases, they may wish to rely
more heavily upon the high spatial resolution of fMRI than the temporal speed of
EEG. This can be particularly valuable when the events targeted for characterization
fall into one of the following categories: (1) events that are largely uncontrolled or
unpredictable (investigations of epileptic activity); (2) events with a large degree of
variability; and (3) events that are invisible to fMRI alone. Serving as an converse
method to fMRI-informed EEG, results obtained by EEG-informed fMRI are similar