I worked a bit learning MNE last Oct-Nov, then had to do other things for a
while. Now I am diving back into MNE, trying to relearn what I did last
fall. One new area I am curious about is how to make a customized source
space. Reading the V2.9 Dec 2007 manual sections 3.5 and 5.4 it seems one
can use the "traditional" or "icosahedral" methods. The latter is clear to
me, but the "traditional" method is not ? what do you mean by traditional
method here?
What I really want is to be able to pre-specify functional areas and use
them as the basic elements of the source space. We are considering
functional networks defined by fMRI experiments, but for the sake of
argument we can here consider Brodmann areas. Can you imagine or suggest a
way that I could use an individual subject's Brodmann areas as the source
space? At one extreme would be assigning a single dipole to each Brodman
area in one hemisphere. At the other would be to model each Brodmann area as
a set of adjacent dipoles uniformly covering it, but linked as a unit.
Well, the later is trivial to do just specify which Brodmann area each
vertex corresponds to and create labels for them (FreeSurfer 5.1.0 does
create estimated labels of some Brodmann areas by default).
To do the Former you could do all the work of the later and then average
/sum (depending on your assumptions) the forward solutions of each
label. You would then need to write your data out to a forward solution
(I don't know of any available code for this) (or just use this info to
calculate everything with your own code).
Let me give some more details, and some real numbers, to make it more
clear what I am after. Let us assume I had an icosahedral cortical surface
model with 10,240 vertices (dipoles). Then assume I have a set of ROIs on
the inflated spherical cortical surface. I with to mask the icosahedral
surface with the ROI mask, and replace each set of dipoles in each ROI
with a single dipole representing that ROI. So if I have a set of ROIs
(Brodmann or otherwise), say 62 in each hemisphere, I would end up with a
total of 124 dipoles to use as my source space, instead of 10,240. In
other works, there would only be 124 unknowns (dipole magnitudes) to
estimate. The position and orientation of each of the 124 dipoles would be
some sort of average position and orientation derived from the set in each
ROI.
I realize this example represents a severe simplification of the cortical
surface, but it could provide some useful information, trading resolution
for a more robust inverse. And we are also going to use a variety of less
severe simplifications; this one is mainly for easy discussion here. A
variant of this idea is to keep all the original 10,240 dipoles, but force
them to vary as a group within each ROI. Again, there would be only 124
unknowns, not 10,240, as all the dipoles in each ROI would be linked to
the same dipole magnitude. In this case, it would be necessary to
calculate the forward solution for each of the original 10,240 dipoles.
Perhaps I do not fully understand the MNE process, but what I am after
does not seem possible by just averaging forward solutions.
As I tried to explain earlier. The easy way to do what you want is
create labels for each area you care about then just extract the data
from each of the labels (as described in the MNE manual) using
10,242/hemi source space. For more details read the MNE manual sections
on labels.
For the option you spoke about earlier where you would calculate out a
dipole for each of the labels the solution is computationally equivalent
to the solution described above i.e. the answer you would get from doing
the above is identical to that you would get from 124 equivalent dipoles
(assuming the 124 dipoles are composed from all of the points in the
source space). The only way you would "improve" source localization (I
do not think calling this an improvement is accurate in most situations.
You are making very strong assumptions (which you may be able to
justify)), would be if you did not use the entire source space i.e.
there were unlabeled areas.
as Dan suggested, it's not a great idea to solve the inverse problem
with a few hundreds of sources. You're likely to under fit your data
and have to mislocated activation peaks. The MNE way is to compute
the estimates on the full source space and then extract the time series
from a label as given automatically by FreeSurfer.