spatio_temporal_src_connectivity failure

[larsoner]

It looks like spatio_temporal_src_connectivity (which calls spatio_temporal_tris_connectivity) doesn't handle some source spaces correctly.

import mne
data_path = mne.datasets.sample.data_path('..')
fname_inv = data_path + '/MEG/sample/sample_audvis-meg-oct-6-meg-inv.fif'
inverse_operator = mne.minimum_norm.read_inverse_operator(fname_inv)
connectivity = mne.spatio_temporal_src_connectivity(inverse_operator['src'], 
                                                    n_times=1)
a = connectivity.shape[0] 
b = sum([s['nuse'] for s in inverse_operator['src']])
assert a == b

a is 8196 while b is 7498, so this fails. I assume this is because it either isn't designed to handle oct source spaces (?), in which case should we have it throw an error? Or is there some way to make it work with oct source spaces?

[larsoner]

Yeah, it looks like it's only designed to work with Ico source spaces, not Oct -- is there some way to make it work with Oct, though? (I'm not too familiar with this...)

[larsoner]

Okay, I take it back, it seems to work okay. If I add patch info and use:

connectivity = mne.spatio_temporal_src_connectivity(src, n_times=1, dist=5e-3)

I get a connectivity matrix of the same size (8196 x 8196). So the inverse operator is selecting a subset of vertices to use...?

[larsoner]

In any case, we should figure out how to deal with the mismatch between the source estimates that are generated, and the size of these connectivity matrices.

[agramfort]

can it be because some sources are too close to the inner skull and excluded from the source space?

[agramfort]

the --mintdist param in the mne_do_forward_solution

[mluessi]

+1 this would make most sense

[larsoner]

Probably. I'm not sure if there is an easy way to detect this just from the source space instance, though, so I'm not sure how we'd throw an error...

[agramfort]

can you give the lines of code that would fail as a test?

[larsoner]

Sure thing:

import mne
data_path = mne.datasets.sample.data_path('..')
fname_inv = data_path + '/MEG/sample/sample_audvis-meg-oct-6-meg-inv.fif'
inverse_operator = mne.minimum_norm.read_inverse_operator(fname_inv)
connectivity = mne.spatio_temporal_src_connectivity(inverse_operator['src'], 
                                                    n_times=1)
a = connectivity.shape[0] 
b = sum([s['nuse'] for s in inverse_operator['src']])
assert a == b

[larsoner]

Basically, a user could assume that the STC they get from using that inverse_operator could be used in spatio-temporal clustering, and it can't very easily since the inverse_operator does not use the whole source space.

[larsoner]

One solution would be to morph the STC from the subject to itself, going from the subset of vertices used by inverse_operator to the full set of source space vertices, but this seems less appealing than somehow being able to generate a connectivity matrix that is compatible with the inverse_operator in the first place.

[agramfort]

Basically, a user could assume that the STC they get from using that inverse_operator could be used in spatio-temporal clustering, and it can't very easily since the inverse_operator does not use the whole source space.

why should the spatio temporal clustering break if sources are ignored
by the inverse operator?
I agree that it can produce artifacts like a clustering not going over
the top of the gyrus but is
that a really issue? just asking.

[larsoner]

The clustering breaks because the STC files generated by that inverse_operator will not have the same number of vertices as the connectivity matrix. They will have b vertices, while the connectivity matrix will have a vertices.

[agramfort]

The clustering breaks because the STC files generated by that inverse_operator will not have the same number of vertices as the connectivity matrix. They will have b vertices, while the connectivity matrix will have a vertices.

how about restricting the connectivity to the used vertices that you
find in the source space structure?

iuse key if I remember.

[larsoner]

That could work. I can give it a shot and see how it goes.

[larsoner]

Alright @agramfort, I think this fixes the issue, and I added a test. I have it raise a warning to the user so they know it's being subsampled.