Unstable results

[ftadel]

Hello,

I'm observing frequently cases where the values are very high for one or two grid points, compared with the other values. An example with 14 electrodes (Emotiv Epoc headset) and the Colin27 template brain in Brainstorm.
Brainstorm protocol: https://www.dropbox.com/s/8qk83zpvwapjulf/bug2_protocol.zip?dl=0
OpenMEEG files: https://www.dropbox.com/s/80r0gsxetb752ts/bug2_openmeeg.zip?dl=0

Are there improvements to expect?
How do you suggest we handle these cases?

Thanks,
Francois

[mclerc]

Hi François
Thanks for reporting this issue.
Brain source number 23153 is apprently causing trouble.
Have not had time to investigate but maybe it's because it is be extremely close to the inner skull surface ?

[ftadel]

Yes, it's probably the case.
I can find a way around this problem by generating a set of layers with different numbers of vertices from Brainstorm. I posted this dataset so that you have a simple example to improve the robustness of your code. It would be great to detect those irregularities automatically, and maybe start with returning an error instead of a wrong result. I think most of our users do not have the background necessary to identify and fix this type of problems by themselves.

The previous issue I posted is more annoying for me, it always crashes no matter how many vertices I use for the different layers...
Francois

[agramfort]

Yes, it's probably the case.
I can find a way around this problem by generating a set of layers with different numbers of vertices from Brainstorm. I posted this dataset so that you have a simple example to improve the robustness of your code. It would be great to detect those irregularities automatically, and maybe start with returning an error instead of a wrong result. I think most of our users do not have the background necessary to identify and fix this type of problems by themselves.

can you use a min distance parameter to ignore dipoles which are too
close and maybe post process with interpolation of the leadfield?

MNE just ignores these points which are too close. Default distance
with the simple BEM solver in MNE is 5mm.

The previous issue I posted is more annoying for me, it always crashes no matter how many vertices I use for the different layers...

no time to look :( :(

[ftadel]

Yes, I will have to add some post-processing of the gain matrices to detect and re-interpolate the aberrant values. But I think it would be more logical if this was done directly in OpenMEEG.

[agramfort]

I agree... we "just" need to find the human resources to do it...

[agramfort]

@eolivi can you have a look at this before the release?