I am working on localizing the frequency-following response (FFR) from subcortical sources using an LCMV beamformer. The source localization results are unexpected. They do not localize to the brainstem or auditory cortex, even though this is an auditory task.
I have a few questions:
How does low SNR affect LCMV beamformer performance and localization accuracy? Because our goal is to derive FFR from infants, the data have very low SNR due to limited recording time.
Is there a recommended way to evaluate the quality or goodness-of-fit of an LCMV beamformer source model? For example, is there similar metrics to the R² (goodness-of-fit) metric used in minimum norm estimates?
I attended the MNE office hours last Friday, and Eric suggested directing these questions to Britta and Marijn. Any guidance or references would be greatly appreciated!!
@tzcheng911 - auditory cortex is difficult to localize with LCMV beamformers because the purpose of this type of beamformer is to filter out highly covariant activity. Since bilateral auditory stimulation produces highly covariant activity in the bilateral auditory cortices, the beamformer can produce odd effects. You may be able to split the channels into left and right channels sets and fit separate beamformers for the left and right cortical hemispheres independently localize the source activity. (It looks like the data in the mne tutorial is single sided auditory stimulation - which might mitigate this source suppression effect).
This paper shows this issue in Fig 2f.
The brainstem localization is difficult because it is a small region and the neurons are not organized parallel to each other (unlike the cortex), so you can have far field cancellation of the magnetic fields. This is also quite distant to the sensors, so you have to contend with the dropoff over distance. It may be possible to localize these, but it is still a difficult target.
beamformers such as LCMV can give completely wrong source estimates if there is any numerical instability when inverting the data covariance matrix. This is the most common problem I experience when using them.
Some things to try:
Set a high value for the reg parameter. Other than making things more blurry, does it also fundamentally change the source locations? If so, there might be a problem with the covariance inversion
Compare the LCMV result with an MNE source estimate. If they agree, then it is less likely that the LCMV is misbehaving.
Thank you all for the reply! @wmvanvliet what would be a reasonable high reg to set? I tried 20 and it still doesn’t change the results. Also may I ask your insights on the two questions in my original post? Thank you!
In my experience, beamformers deal well with low SNR, but there is a breaking point where if the SNR drops below a certain level, the source estimates become wildly inaccurate. Did you try comparing with an MNE source estimate like I suggested?
Hi @tzcheng911
You already received some excellent answers.
I would probably (for testing purposes only!) once set the regularization value to 200% or something. Then you should see blurring for sure.
Apart from that, it is hard to say what might be going wrong, as “unexpected” can mean a lot of things
You can also look at the rank and condition number of your covariance matrix to see if there is anything unexpected there.
Regarding goodness of fit: This is not used for beamforming (at least I don’t know of anything) - beamforming (as opposed to MNE) is adaptive, and expected to suppress (external) noise in your data. Thus, you don’t (always) expect the beamformer solution’s backprojection to match the original data.
