[FieldTrip] Quick eloreta question
r.oostenveld at donders.ru.nl
Thu Oct 25 12:45:03 CEST 2018
Let me CC this to the mailing list, as I think that it might be of iterest to others.
You should preferably be calling eloreta through ft_sourceanalysis. There you can see on line 1016 how it is called for ERP/ERF data with the data covariance (and elsewhere with the cross-spectrum).
The phase information is certainly not ignored: an obvious one that you need is the exact out-of-phase (i.e. 180 degree rotation) of the positive and negative channels that see the same dipole. So taking the “abs” is not appropriate.
But there is sometimes a reason to take “real” and ignore “imag”: For beamforming - where we scan with a single dipole - we do know that the dipole can only be in phase or exactly out-of-phase, which means that the complex part of the CSD cannot be attibuted to the source of interest. That is implemented with the ‘realfilter’ option in beamformer_dics (default=no, consistent with the publication) and beamformer_pcc (default=yes, because it works slightly better). Since we are not doing a scalar beamformer in most cases, but a vector beamformer (which can rotate in 2 or 3 dimensions), it is not obvious under which conditions realfilter works the best. A strongly rotating source has out-of-phase CSD coponents between the different dipole orientations, so in that case realfilter=no should be better able to reconstruct it. But in our data we (anecdotically) tend to get slightly better SNR at the source level for realfilter=yes.
For frequency domain eloreta you are estimating many source simultaneously, where the sources can have any phase relation to each other, so there is a priori no point in only taking the real part of the CSD. Hence it is also not implemented. However, if you were to hypothesize that the underlying source is very one-dimensional (as with an ICA) and not rotating, the same reasoning would apply as for the beamformer and you would expect slightly better performance ignoring the imaginary part of the CSD in the inversion.
PS the getting_started page on “loreta” refers to the specific Loreta software, not the inverse methods. Furthermore, it is >10 years old and probably outdated.
> On 24 Oct 2018, at 19:49, Arnaud Delorme <arno at ucsd.edu> wrote:
> HI Robert,
> Another quick follow up, still on cross-spectrum and Loreta.
> When I use ft_freqanalysis, I get a complex crossspectrum (dataset with 80 trials).
> However, my understanding is that you only use the absolute value of the crossspectrum for source localization (phase information is ignored), and that the crossspectrum should be computed on a single trial basis then averaged the absolute value should be averaged accross trials. It is unclear to me how you can obtain a complex estimate on multiple trials (are you averaging the complex cross-spectrum values across trials - I have done some more test and it seems that this is what you are doing). Attached is a cross-spectrum calculated using this method (custom code, left) versus the absolute value of the cross-spectrum returned by ft_freqanalysis (right) on the same data.
> When I do take the absolute value before doing the average of the cross-spectrum, the eLoreta solution is also more focal.
> <Screen Shot 2018-10-24 at 10.39.10 AM.png>
>> On Oct 24, 2018, at 9:41 AM, Arnaud Delorme <arno at ucsd.edu <mailto:arno at ucsd.edu>> wrote:
>> Hi Robert,
>> I have a quick eLoreta question. In Fieldtrip ft_sourcelocalize, it seems that eLoreta requires the cross-spectrum. I have tried without (or used NaN) but the function is not functioning properly in that case. This means that eLoreta cannot be applied to ERPs, is that correct? What about ICA component scalp topographies (in that case I can weight the cross-spectrum using the channel inverse weight matrix - the cross-spectrum matrix would be proportional to the product of the column in the inverse weight matrix corresponding to the component by its transpose). For spectral decomposition, assuming the spectrum is in the diagonal of the cross-spectrum, is the spectrum field even used at all (I was not able to find information about that).
>> Aso my intuition is that performing statistics (as outlined on this page http://www.fieldtriptoolbox.org/getting_started/loreta <http://www.fieldtriptoolbox.org/getting_started/loreta>) at the voxel level does not make sense if the statistics at the electrode level is not significant.
>> Thank you,
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the fieldtrip