The typical M/EEG workflow¶
Contents
Overview¶
This section describes a typical MEG/EEG workflow, eventually up to source reconstruction. The workflow is summarized in Workflow of the MNE software. References below refer to Python functions and objects.
Preprocessing¶
The following MEG and EEG data preprocessing steps are recommended:
Bad channels in the MEG and EEG data must be identified, see Marking bad channels.
The data has to be filtered to the desired passband.
Artifacts should be suppressed (e.g., using ICA or SSP).
Marking bad channels¶
Sometimes some MEG or EEG channels are not functioning properly for various reasons. These channels should be excluded from analysis by marking them bad as:
>>> raw.info['bads'] = ['MEG2443']
Especially if a channel does not show a signal at all (flat) it is important to exclude it from the analysis, since its noise estimate will be unrealistically low and thus the current estimate calculations will give a strong weight to the zero signal on the flat channels and will essentially vanish. It is also important to exclude noisy channels because they can possibly affect others when signal-space projections or EEG average electrode reference is employed. Noisy bad channels can also adversely affect averaging and noise-covariance matrix estimation by causing unnecessary rejections of epochs.
Recommended ways to identify bad channels are:
Observe the quality of data during data acquisition and make notes of observed malfunctioning channels to your measurement protocol sheet.
View the on-line averages and check the condition of the channels.
Compute preliminary off-line averages with artifact rejection, SSP/ICA, and EEG average electrode reference computation off and check the condition of the channels.
View raw data with
mne.io.Raw.plot()
without SSP/ICA enabled and identify bad channels.
Note
It is strongly recommended that bad channels are identified and marked in the original raw data files. If present in the raw data files, the bad channel selections will be automatically transferred to averaged files, noise-covariance matrices, forward solution files, and inverse operator decompositions.
Artifact suppression¶
SSP¶
The Signal-Space Projection (SSP) is one approach to rejection of external disturbances in software. Unlike many other noise-cancellation approaches, SSP does not require additional reference sensors to record the disturbance fields. Instead, SSP relies on the fact that the magnetic field distributions generated by the sources in the brain have spatial distributions sufficiently different from those generated by external noise sources. Furthermore, it is implicitly assumed that the linear space spanned by the significant external noise patters has a low dimension.
SSP-based rejection is often done using the
mne.preprocessing.compute_proj_ecg()
and
mne.preprocessing.compute_proj_eog()
methods, see
Background on projectors and projections and Repairing artifacts with SSP for more
information.
ICA¶
Many M/EEG signals including biological artifacts reflect non-Gaussian processes. Therefore PCA-based artifact rejection will likely perform worse at separating the signal from noise sources.
ICA-based artifact rejection is done using the mne.preprocessing.ICA
class, see the Background on Independent Component Analysis (ICA) section for more information.
Epoching and evoked data¶
Epoching of raw data is done using events, which define a t=0
for your
data chunks. Event times stamped to the acquisition software can be extracted
using mne.find_events()
:
>>> events = mne.find_events(raw)
The events
array can then be modified, extended, or changed if necessary.
If the original trigger codes and trigger times are correct for the analysis
of interest, mne.Epochs
for the first event type (1
) can be
constructed using:
>>> reject = dict(grad=4000e-13, mag=4e-12, eog=150e-6)
>>> epochs = mne.Epochs(raw, events, event_id=1, tmin=-0.2, tmax=0.5,
>>> proj=True, picks=picks, baseline=(None, 0),
>>> preload=True, reject=reject)
Note
The rejection thresholds (set with argument reject
) are defined
in T / m for gradiometers, T for magnetometers and V for EEG and EOG
channels.
Rejection using annotations¶
The reject keyword of mne.Epochs
is used for rejecting bad epochs
based on peak-to-peak thresholds. Bad segments of data can also be rejected
by marking segments of raw data with annotations. See
Rejecting bad data spans and mne.Annotations
for more .
Once the mne.Epochs
are constructed, they can be averaged to obtain
mne.Evoked
data as:
>>> evoked = epochs.average()
Source localization¶
MNE makes extensive use of the FreeSurfer file structure for analysis.
Before starting data analysis, we recommend setting up the environment
variable SUBJECTS_DIR
(or set it permanently using mne.set_config()
)
to select the directory under which the anatomical MRI data are stored.
This makes it so that the subjects_dir
argument does not need to
be passed to many functions.
Anatomical information¶
Cortical surface reconstruction with FreeSurfer¶
The first processing stage is the creation of various surface reconstructions with FreeSurfer. The recommended FreeSurfer workflow is summarized on the FreeSurfer wiki pages. See also this information FreeSurfer MRI reconstruction.
Setting up the source space¶
This stage consists of the following:
Creating a suitable decimated dipole grid on the white matter surface.
Creating the source space file in fif format.
This is accomplished with using mne.setup_source_space()
and
mne.write_source_spaces()
. These assume that the anatomical MRI processing
has been completed as described in Cortical surface reconstruction with FreeSurfer.
|
Sources per hemisphere |
Source spacing / mm |
Surface area per source / mm2 |
---|---|---|---|
|
1026 |
9.9 |
97 |
|
2562 |
6.2 |
39 |
|
4098 |
4.9 |
24 |
|
10242 |
3.1 |
9.8 |
For example, to create the reconstruction geometry for subject='sample'
with a ~5-mm spacing between the grid points, say:
>>> src = setup_source_space('sample', spacing='oct6')
>>> write_source_spaces('sample-oct6-src.fif', src)
This creates the source spaces and writes them to disk.
Compute Source Space illustrates how the source space is used to compute the forward model.
Creating the BEM model meshes¶
Calculation of the forward solution using the boundary-element model (BEM) requires that the surfaces separating regions of different electrical conductivities are tessellated with suitable surface elements. Our BEM software employs triangular tessellations. Therefore, prerequisites for BEM calculations are the segmentation of the MRI data and the triangulation of the relevant surfaces.
For MEG computations, a reasonably accurate solution can be obtained by using a single-compartment BEM assuming the shape of the intracranial volume. For EEG, the standard model contains the intracranial space, the skull, and the scalp.
At present, no bulletproof method exists for creating the triangulations. Feasible approaches are described in The Boundary Element Model (BEM).
Setting up the head surface triangulation files¶
The segmentation algorithms described in The Boundary Element Model (BEM) produce either FreeSurfer surfaces or triangulation data in text. Before proceeding to the creation of the boundary element model, standard files for FreeSurfer surfaces must be present:
inner_skull.surf contains the inner skull triangulation.
outer_skull.surf contains the outer skull triangulation.
outer_skin.surf contains the head surface triangulation.
Setting up the boundary-element model¶
This stage sets up the subject-dependent data for computing the forward solutions:”
>>> model = make_bem_model('sample')
>>> write_bem_surfaces('sample-5120-5120-5120-bem.fif', model)
Where surfaces
is a list of BEM surfaces that have each been read using
mne.read_surface()
. This step also checks that the input surfaces
are complete and that they are topologically correct, i.e.,
that the surfaces do not intersect and that the surfaces are correctly
ordered (outer skull surface inside the scalp and inner skull surface
inside the outer skull).
This step assigns the conductivity values to the BEM compartments. For the scalp and the brain compartments, the default is 0.3 S/m. The default skull conductivity is 50 times smaller, i.e., 0.006 S/m. Recent publications, see Forward modeling, report a range of skull conductivity ratios ranging from 1:15 (Oostendorp et al., 2000) to 1:25 - 1:50 (Slew et al., 2009, Conçalves et al., 2003). The MNE default ratio 1:50 is based on the typical values reported in (Conçalves et al., 2003), since their approach is based comparison of SEF/SEP measurements in a BEM model. The variability across publications may depend on individual variations but, more importantly, on the precision of the skull compartment segmentation.
Note
To produce single layer BEM models (–homog flag in the C command
line tools) pass a list with one single conductivity value,
e.g. conductivities=[0.3]
.
Using this model, the BEM solution can be computed using
mne.make_bem_solution()
as:
>>> bem_sol = make_bem_solution(model)
>>> write_bem_solution('sample-5120-5120-5120-bem-sol.fif', bem_sol)
After the BEM is set up it is advisable to check that the
BEM model meshes are correctly positioned using e.g.
mne.viz.plot_alignment()
or mne.Report
.
Note
Up to this point all processing stages depend on the anatomical (geometrical) information only and thus remain identical across different MEG studies.
Note
If you use custom head models you might need to set the ico=None
parameter to None
and skip subsampling of the surface.
Aligning coordinate frames¶
The calculation of the forward solution requires knowledge of the relative location and orientation of the MEG/EEG and MRI coordinate systems (see The head and device coordinate systems). The head coordinate frame is defined by identifying the fiducial landmark locations, making the origin and orientation of the head coordinate system slightly user dependent. As a result, it is safest to reestablish the definition of the coordinate transformation computation for each experimental session, i.e., each time when new head digitization data are employed.
The corregistration is stored in -trans.fif
file. If is present,
you can follow Source alignment and coordinate frames to validate its correctness.
If the -trans.fif
is not present or the alignment is not correct
you need to use mne.gui.coregistration()
(or its convenient command line
equivalent mne coreg) to generate it.
Warning
This step is important. If the alignment of the coordinate frames is inaccurate all subsequent processing steps suffer from the error. Therefore, this step should be performed by the person in charge of the study or by a trained technician. Written or photographic documentation of the alignment points employed during the MEG/EEG acquisition can also be helpful.
Computing the forward solution¶
After the MRI-MEG/EEG alignment has been set, the forward
solution, i.e., the magnetic fields and electric
potentials at the measurement sensors and electrodes due to dipole
sources located on the cortex, can be calculated with help of
mne.make_forward_solution()
as:
>>> fwd = make_forward_solution(raw.info, fname_trans, src, bem_sol)
Computing the noise-covariance matrix¶
The MNE software employs an estimate of the noise-covariance matrix to weight the channels correctly in the calculations. The noise-covariance matrix provides information about field and potential patterns representing uninteresting noise sources of either human or environmental origin.
The noise covariance matrix can be calculated in several ways:
Employ the individual epochs during off-line averaging to calculate the full noise covariance matrix. This is the recommended approach for evoked responses, e.g. using
mne.compute_covariance()
:>>> cov = mne.compute_covariance(epochs, method='auto')
Employ empty room data (collected without the subject) to calculate the full noise covariance matrix. This is recommended for analyzing ongoing spontaneous activity. This can be done using
mne.compute_raw_covariance()
as:>>> cov = mne.compute_raw_covariance(raw_erm)
Employ a section of continuous raw data collected in the presence of the subject to calculate the full noise covariance matrix. This is the recommended approach for analyzing epileptic activity. The data used for this purpose should be free of technical artifacts and epileptic activity of interest. The length of the data segment employed should be at least 20 seconds. One can also use a long (
*> 200 s
) segment of data with epileptic spikes present provided that the spikes occur infrequently and that the segment is apparently stationary with respect to background brain activity. This can also usemne.compute_raw_covariance()
.
Calculating the inverse operator¶
The MNE software doesn’t calculate the inverse operator explicitly but rather computes an SVD of a matrix composed of the noise-covariance matrix, the result of the forward calculation, and the source covariance matrix. This approach has the benefit that the regularization parameter (‘SNR’) can be adjusted easily when the final source estimates or dSPMs are computed. For mathematical details of this approach, please consult The minimum-norm current estimates.
This computation stage can be done by using
mne.minimum_norm.make_inverse_operator()
as:
>>> inv = mne.minimum_norm.make_inverse_operator(raw.info, fwd, cov, loose=0.2)
Creating source estimates¶
Once all the preprocessing steps described above have been completed, the inverse operator computed can be applied to the MEG and EEG data as:
>>> stc = mne.minimum_norm.apply_inverse(evoked, inv, lambda2=1. / 9.)
And the results can be viewed as:
>>> stc.plot()
Group analyses¶
Group analysis is facilitated by morphing source estimates, which can be
done e.g., to subject='fsaverage'
as:
>>> morph = mne.compute_source_morph(stc, subject_from='sample', subject_to='fsaverage')
>>> stc_fsaverage = morph.apply(stc)
See Morphing and averaging source estimates for more information.