Corticomuscular coherence (MEG)

[TUTORIAL UNDER DEVELOPMENT: NOT READY FOR PUBLIC USE]

Authors: Raymundo Cassani, Francois Tadel & Sylvain Baillet.

Corticomuscular coherence measures the degree of similarity between electrophysiological signals (MEG, EEG, ECoG sensor traces or source time series, especially over the contralateral motor cortex) and the EMG signal recorded from muscle activity during voluntary movement. This signal similarity is due mainly to the descending communication along corticospinal pathways between primary motor cortex (M1) and muscles. For consistency and reproducibility purposes across major software toolkits, the present tutorial replicates the processing pipeline "Analysis of corticomuscular coherence" by FieldTrip.

Contents

  1. Background
  2. Dataset description
  3. Download and installation
  4. Importing and processing anatomy data
  5. Review the MEG and EMG recordings
  6. Pre-process
  7. Importing data epochs
  8. Coherence 1xN (sensor level)
  9. MEG source modelling
  10. Source estimation
  11. Coherence 1xN (source level)
  12. Coherence 1xN (scout level)
  13. Coherence NxN (scout level)
  14. Script
  15. Additional documentation

Background

Coherence measures the linear relationship between two signals in the frequency domain. Previous studies (Conway et al., 1995, Kilner et al., 2000) have reported corticomuscular coherence effects in the 15–30 Hz range during maintained voluntary contractions.

IMAGE OF EXPERIMENT, SIGNALS and COHERENCE

Dataset description

The dataset comprises MEG recordings (151-channel CTF MEG system) and bipolar EMG recordings (from left and right extensor carpi radialis longus muscles) from one participant who was tasked to lift their hand and exert a constant force against a lever for about 10 seconds. The force was monitored by strain gauges on the lever. The participant performed two blocks of 25 trials using either the left or right wrist. EOG signals were also recorded, which will be useful for detection and attenuation of ocular artifacts. We will analyze the data from the left-wrist trials in the present tutorial. Replicating the pipeline with right-wrist data is a good exercise to do next!

Download and installation

The next sections describe how to import the participant's anatomical data, review raw data, manage event markers, pre-process EMG and MEG signals, epoch and import recordings for further analyzes, with a focus on computing coherence at the sensor (scalp) and source (brain map) levels.

Importing and processing anatomy data

We then need to segment the head tissues to obtain the surfaces required to derive a realistic MEG head model (aka "forward model"). Here, we will perform MRI segmentation with CAT12, this process takes between 30 to 60 minutes.

Once finished, multiple atlases or anatomical parcellations (ICON) will appear in the dataset tree alongside with surfaces for the head (head mask), white matter, cortex (pial envelope) and the midpoint between these last two. The default surfaces are indicated in green. You can display the surfaces by double-clicking on these new nodes. For further information on the anatomy files see the Display the anatomy tutorial.

As part of the MRI segmentation pipeline with CAT12, the anatomy data was normalized in the MNI space, and several anatomical parcellations were computed. These parcellations can be used to create volume and surface scouts, which will be used later in this tutorial to perform the coherence analysis in the source level.

Additional MNI parcellation templates to define anatomical regions of the brain can be used in Brainstorm for MNI-normalized MRI anatomy. See MNI parcellations

Review the MEG and EMG recordings

A new folder is now created in Brainstorm's database explorer and contains:

Display MEG helmet and sensors

Reviewing continuous recordings

Event markers

The colored dots above the data time series indicate event markers (or triggers) saved with this dataset. The trial onset information of the left-wrist and right-wrist trials is saved in an auxiliary channel of the raw data named Stim. To add these markers, these events need to be decoded as follows:

This procedure creates new event markers now shown in the Events section of the tab. along with previous event categories. In this tutorial, we will only use events U1 through U25, which correspond to how each of the 25 left-wrist trials had been encoded in the study. We will now delete other events of no interest, and merge the left trial events under a single event category, for convenience.

These events correspond to the beginning of 10-s trials of left-wrist movements. We will compute coherence over 1-s epochs over the first 8 s of each trial. To that purpose, we will now create extra events to define these epochs.

Pre-process

In this tutorial, we will analyze only the Left trials (left-wrist extensions). In the following sections, we will process only the first 330 s of the recordings, where the left-wrist trials were performed.

Another idiosyncrasy of the present dataset is that the CTF MEG data were saved without the desired 3-rd order gradient compensation for optimal denoising. We will now apply this compensation as follows:

This process creates the SubjectCMC_clean folder that contains a copy of the channel file and a link to the raw file Raw | clean, which points to the original data and to the fact that the 3-rd order gradient compensation will be applied. Brainstorm does not create a physical copy of the actual, large dataset at this stage.

Removal of power line artifacts

We will start with identifying the spectral components of power line contamination of MEG and EMG recordings.

A new raw folder named SubjectCMC_clean_notch is created. Estimate the PSD of these signals to appreciate the effect of the notch filters applied. As above, please remember to indicate a Time window restricted from 0 to 330 s in the options of the PSD process.

EMG pre-processing

Two typical pre-processing steps for EMG consist in high-pass filtering and rectifying.

Two new folders SubjectCMC_clean_notch_high and SubjectCMC_clean_notch_high_abs are added to Brainstorm's database explorer. We can now safely delete folders that are not needed anymore:

MEG pre-processing

We need to remove more artifacts from the MEG traces via the:

  1. Detection and removal of stereotypical artifacts with SSP

  2. Detection of noisy (bad) data segments.

Detection and removal of artifacts with SSP (Signal Space Projection)

Stereotypical artifacts such eye blinks and heartbeats can be identified from their respective characteristic spatial distributions. Their contaminationn of MEG signals can then be attenuated specifically using Signal-Space Projections (SSPs). For more details, consult the dedicated tutorials about the detection and removal of artifacts with SSP. The present tutorial dataset features an EOG channel but no ECG. We will perform only the removal of eye blinks.

Detection of "bad" data segments:

Here we will use the automatic detection of artifacts to identify data segments contaminated by e.g., large eye and head movements and muscle contractions.

We encourage users to review and validate the segments marked using this procedure. In the present case, the segments detected as bad clearly point at contaminated MEG data segments, which we will now label these as "bad".

Importing data epochs

At this point we are finished with the pre-processing of the EMG and MEG recordings. We will now extract and import specific data segments of interest into the Brainstorm database for further derivations. We refer to these segments as epochs or trials. As mentioned previously, we will focus on the Left (wrist) category of events.

A new folder SubjectCMC_clean_notch_high_abs is created for Subject01. It contains a copy of the channel file from the original raw file, and individual trials tagged as Left in a new trial group. Expand the trial group and note there are trials marked with a question mark in a red circle (ICON). These indicate trials that occurred in the bad segments identified in the previous section. All the bad trials are automatically ignored for further processing, whenever dropped into the Process1 and Process2 tabs.

Coherence 1xN (sensor level)

We will now compute the magnitude square coherence (MSC) between the left EMG signal and each of the MEG sensor data.

We can now average magnitude of the MSC across a frequency band of interest (15-20 Hz):

The resulting file mscohere(0.6Hz,555win): EMGlft | tfbands has only one MSC value for each sensor (the MSC average in the 15-20 Hz band). You may visualize the topography of this MSC statistics via 3 possible representations: 2D Sensor cap, 2D Sensor cap and 2D Disk, which are all accessible via a right-click over the MSC node. We clicked on sensor MRC21 below; it is shown in red.

We can observe higher MSC values between the EMG signal and MEG sensor signals over the contralateral set of central sensors in the beta band. Unfortunately, sensor-level connectivity present the disadvantages of: not being interpretable, and being subject to spurious results due to volume conduction.

In the next section we will compute coherence in the source level. To do this, we first need to estimate the sources time series from the sensor data.

MEG source modelling

We will perform source modelling using a distributed model approach for two possible source spaces: the cortex surface and the entire MRI volume. In the surface space, a source grid is made with the source located on the cortical surface obtained from the participant's MRI. For the volume space, the source grid consist of elementary sources uniformly distributed across the entire brain volume. Before estimating the brain sources, we need to derive the sensor noise covariance matrix, and the head model.

Noise covariance

The recommendation for MEG, is to extract basic noise statistics from empty-room recordings. However, when recommended empty-room recordings are not available, as with this tutorial data, resting-state data can be used as proxies for MEG noise covariance. See the noise covariance tutorial for more details.

Head model

The head model, aka forward model, accounts for how neural electrical currents (in a source space) produce magnetic fields captured by sensors outside the head, considering head tissues electromagnetic properties and geometry, independently of actual empirical measurements. As the head model depends on the source space, a distinct head model is required for the surface and volume source spaces. Please refer to the head model tutorial for more in-depth explanations.

Surface

The cortical head model will be derived from each of the 15,000 sources (surface vertices) as defined in the default cortex.

The (ICON) Overlapping spheres (surface) head model now appears in the database explorer.

Volume

The Overlapping spheres (volume) head model is now added to the database explorer. The green color indicates this is the default head model for the current folder (this can be changed by simply double clicking over the head model nodes.)

Source estimation

Now that the noise covariance and head model(s) are available, we will perform source estimation, to find the sources that gave origin to the signals registers in the sensors. From the diverse source estimation methods available in Brainstorm, in this tutorial the minimum-norm imaging method is used. The minimum-norm method estimates the linear combination of the current at each point in the source grid that explains the recorded sensor signals favouring minimum energy (L2-norm) solutions. As result, a large matrix called the imaging kernel is obtained. By multiplying the imaging kernel with the sensor data, it is possible to obtain the estimates of brain sources time series. A different imaging kernel is derived for each of the head models we have produced above: surface and volume. See the source estimation tutorial for more details.

Each dipole in the source grid may point arbitrarily in any direction in a 3D space.

Only for surface grids, the dipole orientation can be fixed to be normal to the cortical surface, this approach is based on anatomical observations of the brain cortex. The result is then in a smaller model that is faster to compute and display.
A discussion on constrained vs unconstrained sources is presented here.

Surface

Here we will estimate the sources in the surface space for constrained (normal to the cortex) and unconstrained dipole orientations.

pro_sources_srfc.png

pro_sources_srfu.png

The inversion kernels (ICON) MN: MEG (surface)(Constr) 2018 and MN: MEG (surface)(Unconstr) 2018 are now available in the database explorer.

Volume

To compute the imaging kernel for the volume source space:

At this point the imaging kernel (ICON) MN: MEG (volume)(Unconstr) 2018 is now also available in the database explorer.

Note that each trial is associated with three source link (ICON) nodes, that correspond to each of the imaging kernels obtained above.

The source link nodes are not files containing the sources time series, instead, the links indicate Brainstorm to: load the corresponding MEG recordings, load the respective inverse kernel, and multiply the two on the fly to generate the sources time series. This approach saves computation time and lot of space on the hard drive.

Coherence 1xN (source level)

Once we have computed the time series for the sources, it is time to compute coherence between the EMG signal and brain sources obtained with each of the imaging kernels.

Let's start with sources from the MN: MEG (surface)(Constr) kernel:

This will create a new tab in the database explorer. This new tab contains only the files that match the search criteria.

Open the Pipeline editor:

pro_coh_srf.png

pro_coh_srf2.png

Do not forget to update the search criteria and the tag to be added to the result.

Coherence with constrained and unconstrained sources

For constrained sources, each vertex in the source grid is associated with ONE time series, as such, when coherence is computed with the EMG signal (also one time series), the result is ONE coherence spectrum per vertex. In other words, for each frequency bin, there is coherence brain map.

In the case of unconstrained sources (surface and volume), each vertex in the grid is associated with THREE time series, each one corresponding to the x, y and z directions; as result, when coherence is computed with the EMG signal, there are THREE coherence spectra, this complicates its representation in the brain, thus, to obtain one coherence spectrum per vertex, the maximum across directions is found for each vertex as at given frequency bin, as shown below.

FIGURE A diagram would help here. (Get it on debug, thus not available on the tutorial)

As an alternative, the x, y and z time series in an unconstrained vertex could be flattened into one time series before computing coherence. See more details at the end of this section.

Results (Surface)

Double-click the 1xN connectivity files for the (surface) source space to show the results on the cortex. If you are not familiar with the option in the cortex figures, check Display: Cortex surface

Find the location and frequency with the highest coherence value.

The highest coherence value is located on the right primary motor cortex (precentral gyrus) at 14.65 Hz for the analysis using constrained and unconstrained sources. Set the the amplitude threshold to 0% to see the extension of the high coherence values.

res_coh_srfc.png

res_coh_srfu.png

MSC @ 14.65Hz (surface)(Constr)

MSC @ 14.65Hz (surface)(Unconstr)

We can see that the results obtained with constrained and unconstrained agree in the location. and it is in agreement with hypothesis and previous results in the literature.

The main difference between the results is that coherence values obtained with unconstrained sources appear smoother, this caused by the maximum aggregation performed across directions (explained in the previous section).

Also the use of Constrained sources avoids the aggregation across dimensions step.

Finally, right-click on any of the cortex figures and select Get coordinates. Then click on the right motor cortex with the crosshair cursor that appears. This coordinates will be useful in the next section.
MRI coordinates X:104.1, Y:111.4 and Z: 223.0

res_get_coordinates.png

Results (Volume)

Double-click the 1xN connectivity file for the (volume) source space. Note that this time the results are shown in the MRI viewer rather than the cortical surface.

res_coh_vol.png

MSC @ 14.65Hz (volume)(Unconstr)

As we can see these results are in line with ones obtained in the surface source space.

Advanced

Flattening unconstrained sources (ADVANCED)

In the previous sections, the dimension reduction across directions in unconstrained sources, aka "flattening", happens on the coherence metric. An alternative approach present in the literature, is to perform the flattening on the time series with methods such as PCA (first component), or the Norm (aka Euclidean norm). This flattening of the time series can be perform in Brainstorm with the process: Sources > Unconstrained to flat map.

A note that the flattened sources are saved as full rather than recordings+kernel. We have tested this flattening approach with simulations and the real data (from this tutorial) and we have found decremental effects on the expected results.

Coherence 1xN (scout level)

So far we computed coherence at the source level, this is to say, a coherence spectrum is compute for each of the 15002 source points. In the case of the unconstrained sources, the are 45006 source points. This large dimensions hinder the further analysis of the results; therefore, the strategy is to reduce the dimensionality of the source space and adopt a parcellation scheme, either in ?surface or ?volume, in other terms we will use the ?scouts. Thus, instead of providing a result (coherence spectrum) per source vertex, the results are aggregated across the sources within scouts. See the scouts tutorial for detail information on how to define and use scouts in Brainstorm.

When computing coherence (or other connectivity metrics) in the scout level, it is necessary to define two parameters: the scout function (mean is often used), and when the within-scout aggregation takes place, this is can be before or after the coherence estimation.

As it can be seen, the After option takes longer and used more resources as it computes the coherence spectrum for each vertex in the scouts, and then, the coherence spectra are aggregated.

BRAINSTORM TEAM
Due to the current implementation of the bst_connectivity, the full source map for each trial (185) are loaded in memory, thus replicate the After, only for the (surface)(Constr) option ~30GB of RAM are needed! (Unconstrained take 3 times that).

Let's here compute the 1xN coherence using scouts, with the Before option. In this tutorial, we will use the ?Schaefer 100 parcelations atlas on the results from constrained sources (surface source space).

Double-click the file. This time it is not displayed on the cortex, but the coherence spectrum for each scout is show. Right click, and it can be also shown as image.

IMAGES

Although the choice of the optimal parcellation scheme for the source space is not easy. The optimal choice is to choose a parcellation based on anatomy, for example the Brodmann parcellation here. In Brainstorm these atlases are imported in Brainstorm as scouts (cortical regions of interest), and saved directly in the surface files as explained in this tutorial here.

Coherence NxN (scout level)

Through analyzing connectivity metrics (coherence in this tutorial), it is possible to derive a connectome.

Here from old tutorial mentioning that source level is not practical cause of the size of the results. Thus a common approach is the use of scouts, as in the previous section. Here we want to write about performing NxN connectivity with the scouts. This should not be done with sources as it leads to a very big matrix that will not fit in the RAM

In our previous examples, the reference signal was 1 dimension, but here it's 3D thus it makes the processing more complex, as there are 9 coherence spectra:

Before: Scout aggregation, 3x3 for each scout pair, then the dimension aggregation.

After: Each source vs each source, thus there are 9 coherence spectra for each source-pair! then the results are aggregated for all the pairs, giving 9 coherence spectra between scouts, then it is aggregated accross diemension.

Here the after is even longer than in the previous sections, thus we will perform only the Before.

Advanced

Script

[TO DO] Once we agree on all the steps above.

Additional documentation

Articles

Tutorials

Forum discussions

[TO DO] Find relevant Forum posts.




Feedback: Comments, bug reports, suggestions, questions
Email address (if you expect an answer):


Tutorials/CorticomuscularCoherence (last edited 2022-02-15 20:12:55 by RaymundoCassani)