[FieldTrip] New NeuroImage

[Schoffelen, J.M. (Jan Mathijs)]

Hi Craig,

Allow me to reply to the list, ’ter lering ende vermaak’.

I was just looking at your new NeuroImage paper comparing toolboxes. Great paper! It left me with a few questions, while clearing up a good deal of them. The code is also really helpful. I was wondering is cfg.lcmv.fixedori = ‘yes’ the only way to specify the orientation for a scalar beamformer?

For the majority of the different beamformer functions in FT, (i.e. lcmv/dics/pcc, in contrast to sam, which behaves a bit different) specifying ‘fixedori’ indeed by default computes a scalar beamformer after ‘optimal’ orientation estimation by means of an svd on the vector beamformer output. Yet, in combination with fixedori=‘yes’, you can also specify the weightnorm option, which defaults to ’no’, but which can also be specified to be ‘arraygain’/‘unitnoisegain’/’nai’. These settings result in different heuristics for the ‘optimal’ orientation estimation, as per (e.g.) the descriptions in the Sekihara/Nagarajan book.

I think I remember that in Brainstorm you can specify this orientation, which I suppose would be best specified as perpendicular to the cortical sheet. Can FT do this, or does it always project the 3 axes to the largest axis (SVD, etc.)?

If you want to use a fixed orientation that for instance is based on a cortical orientation constraint, you’d need to plug that in yourself. That would best be done at the step at which you compute the leadfields (I assume you use the recommended two-step approach, in which you first compute the forward solution as per ft_prepare_leadfield, and plug its output into the cfg for ft_sourceanalysis). I think the option to use there would be to have a mom (or ori, please check the documentation to be sure)-field associated with the source positions for which you want to compute the leadfield. I think that the sourcemodel.mom should be of dimensionality 3xnpos, but I am not 100% sure.
As a side note with respect to cortical orientation constraints: 1) I am not a big fan of this, since it assumes (and relies on) a very good coregistration between the anatomical information and the MEG sensor-array, as well as on the absence of appreciable subject motion throughout the measurement. 2) the quality of the estimates of the local ’orthogonal’ directions probably highly depends on the local folding pattern of the cortical surface, in interaction with the granularity of the mesh.



The other question is: I haven’t been able to find any info on reduce rank. So here it seems we can throw out the smallest orientation. This means X,Y,Z of the moment? Why are we doing this? Why does it help?

Indeed, by default for MEG, FT computes the leadfield of a dipole that lives in a plane, where the most radial orientation is not represented . This is to avoid the beamformer spatial filter to potentially project a lot of noise onto this orientation. Or, put differently, since the most radial orientation of the leadfield has a very small gain, the component of the spatial filter that corresponds to this orientation can blow up in your face, leading to ugly results. In that case we’d rather be safe than sorry, and thus remove this source of mayhem from the forward solution.

Best wishes,
JM




Best,

C.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20200716/3622319f/attachment.htm>