I worked through creating a discrete source space of the right amygdala and
performing dSPM source reconstruction in the SPM dataset. My code is
available on my GSoC blog.
I've attached a copy of the results for data filtered between 1-45 hz. The
grey region indicates statistical significance using an independent t-test
with fdr correction.
I would greatly appreciate any feedback you could offer.
I have a few comments to make on the code so a PR on github would allow
me to comment inline.
Regarding the approach, it does what it is supposed to do in a pretty
elegant way, but
it is dangerous to only run the inverse method on such a tiny
source space. An activation from the cortex would necessarily be projected
to the amygdala. My hunch, is that an LCMV beamformer would suffer
less from this.
I think the next step would be to be able to combine a surface source space
and the deep volume source space you create.
my 2 cents, if you look at the waveform carefully, the responses follows
the same time course as the cortical responses. Putting this example will
give very wrong impression about MEG and deep sources.
Putting premature examples might not be great idea
It also might be a reasonable time to discuss a bit of the physics and physiology of what MEG is and is not sensitive to. MEG is not particularly sensitive to four things:
1. Radial sources
2. Deep sources
3. Sources where the neurons are not aligned
4. Even worse are sources with randomly aligned neurons that are relatively spherical (and small)
The amygdala strikes out on the bottom three. Thus, in my opinion, it is fairly unlikely that MEG can pick up activity from the amygdala. I would love to hear other opinions.
Best wishes,
Avniel
Avniel Ghuman, Ph.D.
Director of MEG Research
Assistant Professor of Neurological Surgery, Neurobiology, and Psychiatry
Faculty in the Center for the Neural Basis of Cognition
University of Pittsburgh
I would add that the issue of having the software machinery to potentially try localizing deep sources is a separate issue from trusting the obtained solution..Thus, for the purposes of demos and code development, localizing simulated deep source activity might be the best scenario where the ground truth is known..
If real data demos are really desired, something like the ABR might be a potentially safer candidate...
I agree with that have been said. We're at the beginning
of the GSOC and our first objective is to have the machinery
working. Then we'll do the science right.
@sheraz : When I said we need the cortical surface in the source
space to not just estimate the sources on the amygdala it
is precisely because I had in mind the issue you raised.
Alan please continue to assemble the machinery, maybe next time
avoid the stats at the end ( ) as it is certainly not a good
conclusion.
cheers,
Alex
PS : with all this interest on this topic, it looks like we're going
to have a lot of early adopters to test
Thanks everyone for the feedback. I've put up the current code as a pull
request, with the understanding that there will be major changes before its
merged.
