110
Y. Zhang
to results from MRI alone; they are largely static and comprised of a small number
maps (or a single map) of activated 3D voxels. As with fMRI, results are not subject
to the volume conduction problem or intense calculation and are not susceptible to
the error that may arise from EEG source localization. While the lack of temporal
resolution may lead researchers to overlook EEG-informed fMRI when attempting to
characterize cortical activity, the robust nature of the MRI results and guiding effects
of EEG have made this a clinically valuable and worthwhile imaging approach.
When attempting to perform EEG-informed fMRI, we must recall that our primary
data method of choice is fMRI. Whereas fMRI-informed EEG built results from the
fMRI into the EEG source localization framework via the source covariance matrix,
we will turn once more to the General Linear model presented in the section on fMRI
analysis. Recall the model:
y xβ + e
where y is a t × 1 time-course of bold activity for every voxel within the MRI volume, x is an r × t matrix representing the timecourse, t, of the r regressors within
the model, β is the r × 1 matrix of static regressor coefficients, and e is the t × 1
matrix representing the error of the system at any given time point. Recall further
that the r regressors used within the model are a series of t × 1 signals used to model
the status of the various factors that can directly impact voxel intensity—motion,
drift, noise, stimuli of interest, etc.—which are ultimately convolved with general or subject-specific models of the cortical hemodynamic response. Logically,
an EEG signal of the activity of interest can be incorporated directly into this
model as one of these regressors. Integration of this type may require some extra
steps during this processing, as EEG signals would need to be downsampled to
match the timescale of collected fMRI. Once appropriate processing has been performed, however, this incorporation becomes straightforward and requires only
that the EEG data of interest be added to the x matrix as a regressor of interest
for the fMRI statistical analysis. More pressing questions, then, are exactly what
EEG feature is selected and how that selection is performed-questions that will fall
to individual researchers.
5.3.2.3 Symmetrical Methods
At this point, we have discussed the predominate asymmetrical methods for combining EEG and fMRI. These methods have been used broadly and have significantly
impacted the field of biomedical imaging. Reviewing these applications and their
biases, however, it becomes clear that the multimodal integration is only implemented as an extension of existing unimodal methods. It is natural that continued
development would focus instead on novel methods that better utilize both methods.
The inherent differences between EEG and fMRI have made this advancement difficult, however, requiring more mathematically advanced methods to achieve these
tasks. Unlike the EEG-informed fMRI and fMRI-constrained EEG, which each have
Y. Zhang
to results from MRI alone; they are largely static and comprised of a small number
maps (or a single map) of activated 3D voxels. As with fMRI, results are not subject
to the volume conduction problem or intense calculation and are not susceptible to
the error that may arise from EEG source localization. While the lack of temporal
resolution may lead researchers to overlook EEG-informed fMRI when attempting to
characterize cortical activity, the robust nature of the MRI results and guiding effects
of EEG have made this a clinically valuable and worthwhile imaging approach.
When attempting to perform EEG-informed fMRI, we must recall that our primary
data method of choice is fMRI. Whereas fMRI-informed EEG built results from the
fMRI into the EEG source localization framework via the source covariance matrix,
we will turn once more to the General Linear model presented in the section on fMRI
analysis. Recall the model:
y xβ + e
where y is a t × 1 time-course of bold activity for every voxel within the MRI volume, x is an r × t matrix representing the timecourse, t, of the r regressors within
the model, β is the r × 1 matrix of static regressor coefficients, and e is the t × 1
matrix representing the error of the system at any given time point. Recall further
that the r regressors used within the model are a series of t × 1 signals used to model
the status of the various factors that can directly impact voxel intensity—motion,
drift, noise, stimuli of interest, etc.—which are ultimately convolved with general or subject-specific models of the cortical hemodynamic response. Logically,
an EEG signal of the activity of interest can be incorporated directly into this
model as one of these regressors. Integration of this type may require some extra
steps during this processing, as EEG signals would need to be downsampled to
match the timescale of collected fMRI. Once appropriate processing has been performed, however, this incorporation becomes straightforward and requires only
that the EEG data of interest be added to the x matrix as a regressor of interest
for the fMRI statistical analysis. More pressing questions, then, are exactly what
EEG feature is selected and how that selection is performed-questions that will fall
to individual researchers.
5.3.2.3 Symmetrical Methods
At this point, we have discussed the predominate asymmetrical methods for combining EEG and fMRI. These methods have been used broadly and have significantly
impacted the field of biomedical imaging. Reviewing these applications and their
biases, however, it becomes clear that the multimodal integration is only implemented as an extension of existing unimodal methods. It is natural that continued
development would focus instead on novel methods that better utilize both methods.
The inherent differences between EEG and fMRI have made this advancement difficult, however, requiring more mathematically advanced methods to achieve these
tasks. Unlike the EEG-informed fMRI and fMRI-constrained EEG, which each have