source localization given induced spectra

Thu Nov 6 19:29:05 CET 2008

Dear Jan-Mathijs,

Lovely suggestion. This works fine, although I must intervene between
the freqanalysis and dipole fit to put the information (I'm using the
real part of the cross spectrum) in the correct form. The automatic
route calls freq2timelock, which does indeed accept cross spectra, but
prepare_freq_matrices seems to expect them to be square, because it
chokes on such a 1-d cross spectrum. It seems more geared towards
beamformer fitting, which I assume requires the full CSD matrix. I
need to look at this further, and may have another question, but just
to let you know now thanks for the suggestion.

John

On Nov 5, 2008, at 1:52 AM, jan-mathijs schoffelen wrote:

> Dear John,
>
> Probably I was a bit rash in answering to your earlier mail and did
> not think it through completely. Apologies for that and for any
> confusion created.
> Let's give it another try:
> As you pointed out, induced power can show a somewhat dipolar
> topography, but you cannot fit a dipole to this because of the fact
> that it is positive all over the place. However, if you would indeed
> have some phase information in your topography, showing the ingoing
> and outgoing field (MEG), or positive and negative potential (EEG),
> one could fit a dipole. I was talking rubbish when suggesting to use
> fourierspectra, because it is forbidden to average them across
> trials (not actually forbidden, but it does not make sense).
> However, it would make sense to average cross-spectral densities
> across trials (not the csd's between all channel combinations, but
> between a well chosen set of channel pairs). Why would this be the
> case? Well, if you cleverly choose your reference signal (I suggest
> one of the strong blobs in your dipolar powerspectrum), and you
> compute the cross-spectral density between this guy and the rest of
> the channels, then you can do business because the csd's represent
> the estimated phase-difference between the reference signal and the
> rest. The interesting signal (your dipolar field) which is buried in
> your single trial data now has a phase of 0 (channels are in the
> same blob of the dipolar field) or 180 degrees (channels are in the
> opposite blob of the dipolar field) with respect to the reference,
> and importantly this is the same for all trials. This means that you
> can average the cross-spectral densities across trials to generate a
> complex-valued topography. Plotting the real-part (or imaginary
> part) of this should lead to a nice dipolar pattern with positive,
> and negative values.
> As far as I know dipolefitting can deal with cross-spectra as an
> input, so my revised approach would be: compute spectrograms and
> identify your time-frequency region of interest. Then plot the
> topography and compute a sensible reference channel. Then call
> freqanalysis again, but with output='powandcsd', and channelcmb =
> {'all' 'yourchosenchannel'}. Then try a dipole fit.
>
> Hope this helps,
>
> Jan-Mathijs
>
>
> On Nov 4, 2008, at 5:02 PM, John Iversen wrote:
>
>> Dear Jan-Mathijs,
>>
>> Thanks for the quick reply.  I'm sorry to hear I was right :)
>>
>> I may misunderstand what you've suggested, but is it not the case
>> that if I chose cfg.output='fourier' it will average fourier
>> spectra across trials (cares about phase) instead of power spectrum
>> (phase blind, as I had been doing). In the end wont I simply get
>> the equivalent of the fourier spectrum of the timelock average,
>> which is substantially different from the induced spectrum that
>> interests me?
>>
>> In many cases the sensor topography of the induced power looks
>> somewhat dipolar, with two power peaks, so I may well try to do a
>> fit, but optimizing not on the field but the field power (this
>> would require modifying the output of the forward model within
>> dipolefitting)--it will not be able to get the polarity of the
>> dipole, but should be able to get a location. Maybe? It seems
>> possible in principle, but I wonder if anyone has practical
>> experience with this.
>>
>> I feel there should be a way to study this kind of question!
>>
>> Best,
>>
>> John
>>
>> On Nov 3, 2008, at 1:18 PM, jan-mathijs schoffelen wrote:
>>
>>> Dear John,
>>>
>>> No, it is not possible to perform source localization on the
>>> spectrograms as you define them. You quite rightly point out that
>>> a spatial topography of power is always positive, so cannot
>>> account for a proper dipolar pattern. However, source localization
>>> of induced changes in oscillatory activity is possible. There's
>>> actually a nice tutorial on the fieldtrip website: "localizing
>>> oscillator sources, using beamformer techniques".
>>> Alternatively, you can actually call dipolefitting using frequency
>>> data as an input, but this requires either fourier-data, or cross-
>>> spectral densities between all channel combinations. Usually the
>>> fourier-data is more memory efficient. In this case I would
>>> propose a two-step strategy: compute spectrograms to identify your
>>> time-frequency region(s) of interest. Then call freqanalysis
>>> again, with cfg.output = 'fourier'. Then I would guess that
>>> dipolefitting runs through... At least it's worth a try.
>>>
>>> Yours,
>>>
>>> Jan-Mathijs
>>>
>>>
>>> On Nov 3, 2008, at 5:26 PM, John Iversen wrote:
>>>
>>>> Hello,
>>>>
>>>> Is there a way to do source localization on induced spectrograms?
>>>> (Induced spectra being the mean of individual trials' power
>>>> spectra.) Conceptually I am not sure how this would work, given
>>>> that one starts with topographies of real, positive-valued power,
>>>> with no phase information, so any dipole fit could be at best
>>>> sign-indeterminate.There is no facility within fieldtrip to do
>>>> such a thing as far as I can tell (induced spectra were
>>>> calculated freqanalysis on multi-trial data and are within
>>>> the .powspctrm field of the result, which is not handled by
>>>> freq2timelock, and thus cannot feed any of the localization
>>>> routines).
>>>>
>>>> What is of actual interest are task-related fluctuations of the
>>>> power around a (much larger, and topographically varied)
>>>> baseline. Is there a way to say where in the brain are the
>>>> (presumed) subset of neural sources that vary in power with time?
>>>>
>>>> Thanks,
>>>>
>>>> John
>>>>
>>>> ----------------------------------
>>>> 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.
>>>
>>> ----------------------------------
>>> 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.
>>
>> ----------------------------------
>> 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.
>
> ----------------------------------
> 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.

----------------------------------
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.