Seemingly empty lead field vector maps when using Deep Brain Analysis

Discussions
1

Hi Brainstorm team,

I am attempting to estimate the deep sources of our sleep EEG data, but am running into an issue I cannot solve and am hoping you can help me. This is new territory for me, so I apologize in advance if there is a basic answer to my issue that I am missing.

For starters, the data: I have T1 anatomical scans for each individual that have been recon-all'ed in FreeSurfer, and the EEG are preprocessed in our homegrown MATLAB software and imported into Brainstorm as single-timepoint scalp maps for each individual (two identical time samples per map actually, as BST did not like importing a single sample FieldTrip-format dataset). The cortical and subcortical surfaces from FreeSurfer look appropriate after import into BST, to my eye at least, as do the EEG scalp maps and electrode locations. I generated realistic mixed (cortical/subcortical) forward models using OpenMEEG as described in the DBA and other tutorials (thank you by the way, for having such fully featured tutorials), then estimated sources using MNE/sLORETA with the identity matrix for noise covariance.

The issue I am having is that for some individuals, the lead field vector maps seem to appear empty when I quality control the OpenMEEG output with the "View lead field vectors" menu option. Further, for these individuals the source estimates have extremely small magnitude, to the point of labelling the color axis with "invalid scale", and seemingly noise-driven widespread diffuse activation compared to other participants. This is not the case for all individuals though, as some data points have lead field vector maps that seem sensible to me- meaning that there are small dipoles throughout the cortex and subcortical structures, with larger vector magnitudes at cortical patches directly under the electrode site. The individuals with sensible seeming vector maps also have more sensible source estimates, with activations centered in cortical patches that I would a priori theoretically expect to see. Additionally, for the problem participants, using OpenMEEG to generate a forward model based only on their cortex (i.e., not the mixed surface with subcortical structures added) results in a more sensible seeming lead field vector map, which leads me to suspect something is going wrong (or more accurately, that I am doing something wrong) with the mixed model/deep source estimation.

Can you all offer any guidance on where to look for what might be going wrong for me? Or, is this the nature of the deep source modeling, that it does not work well for some anatomies? The hippocampal, cerebellar, and striatal contributions are of primary importance to our research question, so if there is any way to make this work I would like to find it. Moreover, performing a separate cortex-only analysis of these data based on the template ICBM152 brain (for troubleshooting) reveals reasonable significant source-domain group effects in our comparisons, so I do believe there must be room for improvement in my OpenMEEG/DBA workflow.

In case it is useful, here is a more detailed step-by-step of the procedures I have taken so far per subject in BST:

  1. Imported FreeSurfer recon-all folder and preprocessed FieldTrip-format EEG dataset

  2. Created a vertex-downsampled atlas from the imported subcortical structures containing hippocampus, cerebellum, brainstem, caudate, putamen, pallidum, thalamus, amygdala, and accumbens, merged with cortex into a mixed surface

  3. Created a source modeling atlas for the mixed surface, and set all structures to deep brain modeling

  4. Generated Scalp, Inner Skull, and Outer Skull BEM surfaces with 1922 vertices each

  5. Used "Force inside skull" menu option to correct cerebellum and brainstem clipping outside skull (this was in response to warnings that vertices were outside the skull for most participants)

  6. Edited the MRI registration of electrode locations using the "Project electrodes on surface" function (we do not have individually measured electrode locations e.g. Polhemus, only standard cap layout positions)

  7. Computed the head model using OpenMEEG BEM with adaptive integration and all other default settings

  8. Set the noise covariance to be the identity matrix

  9. Estimated sources with MNE/sLORETA

If additional information or any screenshots would be helpful for diagnosis please let me know, and I will gladly provide them. Thank you in advance for any advice you can offer, and thanks again for your work on Brainstorm!

  • Ahren
2

The issue I am having is that for some individuals, the lead field vector maps seem to appear empty when I quality control the OpenMEEG output with the "View lead field vectors" menu option.

This could be only a normalization issue in the display.
Press H for help with the display options.

Set the noise covariance to be the identity matrix

Whenever possible, provide a noise covariance computed from the subjects recordings. This is by far where you can expect the most improvement in your analysis. The recordings used to compute the noise covariance should be pre-processed exactly as the EEG data for which you are trying to localize sources.
https://neuroimage.usc.edu/brainstorm/Tutorials/NoiseCovariance#Variations_on_how_to_estimate_sample_noise_covariance

Can you all offer any guidance on where to look for what might be going wrong for me? Or, is this the nature of the deep source modeling, that it does not work well for some anatomies? The hippocampal, cerebellar, and striatal contributions are of primary importance to our research question, so if there is any way to make this work I would like to find it.

These mixed head models were designed primarily for single subject MEG recordings.
The type of data they produce is very difficult to deal with, especially for any post-processing task, and very difficult to debug. One problem in your case is that they do not offer any accurate way of template registration for group analysis: you would not have any good solution for projecting the source maps on the template brain for group analysis.
With EEG source estimation, you can't really expect to have a spatial resolution higher than a lobe-level. Any deep source, if visible on the scalp EEG, would be visible almost on all the electrodes, therefore difficult to localize... If you are really confident you can observe some of your deep regions of interest on the continuous surface EEG, maybe an ICA approach would be better indicated than a minimum norm solution.
Working with non-averaged data is also a challenge. For being visible on the scalp EEG, an hippocampal source should be averaged over tens or if possible hundreds of trials.

I'm sorry I don't know how to help you fix your current issues. Instead I can encourage you to simplify your analysis, with one of these options: