use of dipoles as spatial filters (for SEP)

jan-mathijs schoffelen j.schoffelen at PSY.GLA.AC.UK
Wed May 28 16:15:54 CEST 2008


Hi Markus,

Did you consider trying out the beamforming algorithm with multiple
dipoles as a source model?
As opposed to the dipole-fitting (which pinv'es the concatenated
leadfields), the beamformer actually tries to suppress contributions
of the other guys (provided they are not too much correlated in
time). Obviously, the suppression will be more robust against
violations of the temporal uncorrelation-assumption, when the
leadfield-columns are less correlated. However, I would guess it's
worth a try, and it's exactly what you try to achieve: to avoid
mixing up correlated leadfield components. But perhaps I did not
grasp the problem completely... ;o)

Yours,

JM

On May 28, 2008, at 4:00 PM, Markus Bauer wrote:

>
> Hi fieldtrippers (in particular probably Robert)
>
> I have a few slightly more advanced questions regarding the use of
> dipoles as spatial filters and the creation of multi-dipole models
> with fieldtrip (questions in bold)
> I know that the functionality of fieldtrip is not designed for
> multi-dipole models, but neither Beamformer nor MNE worked somewhat
> reliably (only worked in rather few subjects and then was usually
> pretty smeared still), as often is the case...(for evoked fields in
> this case). So I finally decided to once go back to the "gold
> standard" of somatosensory EEG/MEG research
>
> I have used a serial fit-procedure to fit components of the SEP to
> EEG data. The early components (50 and 80 ms post response) are
> largely determined by two single dipoles, plus a symmetric dipole
> later for bilateral S2. So it is kind of ok to fit simple dipoles
> to those (though clearly not ideal)
>
> My first question:
> Are there plans to extend the functionality of dipolefitting to
> allow for proper serial fit (i.e. adding to be fitted dipoles to
> existing ones) ? I guess it would not be that much rewriting to
> include a fixed dipole and only fit the parameters of the other?
>
> Anyway, in order to create spatial filters I have then fitted the
> orientation of each dipole (estimated from the moments) and thereby
> reduced the leadfield to a onedimensional one.
> In order to get to the sourcewaveforms, I have concatenated the
> (one dimensional) leadfields associated with each source into a
> common leadfield matrix (4 sources corresponding to 4 rows and 124
> columns for each channel) and then took the pseudoinverse of this
> combined leadfield matrix - for use as spatial filters to obtain
> sourcewaveforms - for the "complete solution".
>
> When I looked at those filters obtained, however, I realized that
> it did not work that nicely, since the filters looked for at least
> two components very similar - more similar than the respective
> leadfields!! (which is contra-intuitive)
> The topographies of the peaks where I fitted dipoles to are
> beautiful single dipolar topographies and so are the single
> leadfield matrices, but the filters are anything else than clean.
> It is clear that they will represent a mixture between the sources
> - mutually "supressing" each other. But at least for two of them
> (the early sources for 50 and 80 ms with clearly distinct topos,
> presumably area3b and area1) they pretty much mixed up. I was
> surprised about this cause the topoghraphies look so distinct and I
> would have assumed this should come out nicely (also considering
> all the Hari-studies)
>
> Anyway, a look at the correlation matrix of the leadfields showed
> that despite the topographies of those dipoles had quite a
> different orientation they were correlated at 0.77 (weights over
> channels). So I guess what happens is identical to a suppressor
> effect in multiple regresion - when the predictors are too
> correlated their weights get mixed up....
>
> I tend to use the individual spatial filters now - not the
> pseudoinverse of the combined source model (consisting of only
> four, well identifiable sources).
> My question:
>
> Is this justified - to take individual filters when having a multi-
> dipole model ?
>
> I also realized that when I only included the first two sources
> (the ones described above), and take the pseudoinverse the
> separation succeeds much better.
>
> - Might prior regularization of the leadfield help ?
> - Are there any standards about this - when to still use a combined
> leadfield model ? I guess should be the same as in multiple
> regression cause formulas more or less identical...? But I guess
> nobody publishing sourcewaveforms from multidipole model cares
> about it...
>
> I was a bit "shocked" to see how much this can falsify the results
> (the waveforms), I also wouldn't have expected the leadfields (and
> filters...) of such rather different sources to be that strongly
> correlated. I assume it will be not much different in MEG, cause
> they were both tangential dipoles, looking almost identical as in
> MEG (apart from 90deg rotation due to field geometry).
> My strategy is therefore to either use individual filters of single
> dipoles or reduce the dipolar model...any other suggestions ??
> It seems it is really a very ill-posed problem that we're spending
> our time with....
>
>
> thanks and best wishes
> Markus
> ----------------------------------
>
> The aim of this list is to facilitate the discussion between users
> of the FieldTrip toolbox, to share experiences and to discuss new
> ideas for MEG and EEG analysis.
>
> http://listserv.surfnet.nl/archives/fieldtrip.html
>
> http://www.ru.nl/fcdonders/fieldtrip/
>


----------------------------------
The aim of this list is to facilitate the discussion between users of the FieldTrip  toolbox, to share experiences and to discuss new ideas for MEG and EEG analysis. See also http://listserv.surfnet.nl/archives/fieldtrip.html and http://www.ru.nl/fcdonders/fieldtrip.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20080528/8b60315f/attachment.html>


More information about the fieldtrip mailing list