Source localization
Last updated
Last updated
The procedure of localizing brain activity while knowing the electric (EEG) or magnetic signals (MEG) on or over the scalp is called inverse solution.
According to the ASCT approach, reconstruction of cortical activations is performed with prior localization of ICs using the MNE (minimum norm) method. After localizing ICs based on their topographies, activation of brain voxels can be calculated, accounting for the spatial extent of ICs and their time-courses.
Before starting, we recommend some theoretical introduction to the ionverse solution methods. This can be found elsewhere, but our publication: [in review] will provide an accessibly written introduction.
The source-reconstruction settings are grouped in the loc section of ExSetup. For reliable localization of cortical activity, a proper choice of λ regularization parameter is a crucial issue. However, determining its optimal value can be challenging. ASCT offers two possibilities for setting up λ. When loc.norm is set to 'lambda', it will use a fixed value from the loc.lambda parameter (a rough guess to start with would be 0.1) .
Here, we recommend a more reliable dynamic estimation of λ based on proprietary noise values that can be predetermined for a particular measurement environment together with IC signal parameters. In this case, loc.norm should be set to 'cov' and the system noise level provided with loc.noiselev. Empirical selection of this latter parameter is described below.
Additionally, to prevent the bias towards the superficial sources, which is present when the classic MNE is used, the depth weighting can be enabled with the loc.depthWeight parameter (recommended value 0.5, which is also a default).
The noise level parameter used for estimating λ accounts for the noise in the MEG/EEG system. It is possible to determine its value by examining the quality of source localization performed on clean pilot data with different settings. This value can be considered specific for a given device and can be applied across different measurements in the same measurement environment.
The prerequisite for this procedure is to select some 'good' ICs of brain origin that have consistent, smooth topographic images (saved in 5_ICA2/[dsname] as _BRAIN_BEST_ITER_X_MAPS.jpg file, where X is the ICA2 run selected as the best decomposition). It is recommended to select one of the subject's first ICs due to their relatively high SNR. With loc.norm='cov' setting, a series of IC localizations with different noise settings can be performed automatically using the sc_lambda_est script. Examining the output images in7_SOURCEACT/icloc/lambdafigs will help to determine the optimal value of the loc.noiseLev parameter. This value would produce a possibly focal but also unitary solution. Too high values will result in too blurred, i.e., spatially extended solutions, while too low values will make them fragmented and scattered in space. For an example, see the figure below.
To localize cortical sources, the propagation of an electric / magnetic field in the head has to be determined. To this end, a proper head model is required, which is a geometrical representation of head tissues with their electric or magnetic properties (conductivity for EEG or magnetic permeability for MEG). With the proper head model, the leadfield (or gain) matrix can be calculated. This matrix tells how strong activity of cortical voxels can be picked up by measuring sensors. On the other hand, the spatially extensive activity of cortical voxels is modeled by a dense net of cortical source dipoles called a source model. Only having these three puzzles available—the head model, the source model, and the leadfield—the estimation of cortical activation based on sensor signals becomes possible.
If individual MRI images are not accessible, SPM8 template (averaged) head and source models can be used instead at the cost of decreased accuracy. By default, SPM8 templates are used. Other template models can also be used (for example, for children's subjects), which need to be specified by the rc.headTemplate and rc.sourceTemplate parameters.
MEG sensor locations are by default read from the raw data header or can be provided with XXXXX parameter. For EEG measurement, even more accuracy can be achieved with individual electrode positioning that can be obtained from 3D modeling devices. The rc.elecRealPos parameter has to point at the directory with individual electrode mat files containing locations of sensors and fiducials in the same coordinate system saved in the FieldTrip format.
To improve the accuracy of source localization, individual head- and source models are necessary to account for differences in head/brain shapes. Custom head models can be individually constructed from T1 MRI images. Relevant settings include T1 files location and format (rc.mriRawFolder, rc.mriExt, rc.mriCoordsys) and further loc.headmodelMethod. We recommend realistic 5-layer SimBio head models for EEG and a realistic homogenous head model for MEG, which are chosen by default. An additional shell script is provided for batch preparation of individual source models based on FreeSurfer cortical reconstruction and HCP-workbench dipole net downsampling (/tools/FS-recon.sh) according to the procedure described in the relevant FieldTrip manual: .
The for more detailed instruction on 3D electrode scanning and preparing sensor location files see:
