[FieldTrip] using eeglab's ICA data with beamformer
johanna.zumer at gmail.com
Mon Mar 17 12:34:51 CET 2014
Beamforming is one type of inverse method and it requires the sensor data
covariance matrix. It also assumes that underlying sources are
uncorrelated. If you try to input data from only 1 IC, all the sensors
will covary together and any underlying sources that contribute to this 1
IC will be 100% temporally correlated.
You can apply beamforming to a mixture of ICs, mixed back from the
'component space' back to the 'sensor space' (see ft_rejectcomponent).
This is often done for example by excluding a few components which are
artefactual and keeping many remaining (presumably brain) components.
Normally the 'rank' of your covariance matrix is equal to the number of
sensors (rank is the number of independent rows or columns, i.e.
contributions), and so by removing a few ICs from the data, you have
reduced the rank. If your rank is reduced relative to number of sensors,
then you should use the regularisation option (cfg.lambda). This is true
for either LCMV or DICS. However, this is usually only done when the rank
is still well above 1.
However, if you wish to localise the underlying sources of just 1 (or a
small number of mixed) ICs, then better to use a min-norm method which does
not depend on the sensor covariance. In this case, you still must
re-project the IC back to sensor space in FieldTrip for the labels of the
data to match the labels of the leadfield.
http://fieldtrip.fcdonders.nl/tutorial/shared/virtual_sensors (<- LCMV)
Hope this helps; please ask again if you need further clarification, or
please give a specific error message.
2014-03-17 0:21 GMT+00:00 Tyler Grummett <tyler.grummett at flinders.edu.au>:
> Hello fieldtrip,
> I am new to the fieltrip toolbox and Ive come from eeglab to do some
> beamformer and connectivity processing.
> Ive used the function eeglab2fieldtrip, with one of the arguments being
> 'componentanalysis'. I tried using the data in the beamformer tutorial and
> everything worked fine
> until I got to the ft_sourceanalysis script where it used the channel
> labels and positions from the original non-ICAed data. So the code crashed
> because the ICA labels
> and channels labels didnt match.
> I was wondering whether beamformer can be used with ICs from eeglab, and
> if so, where I am going wrong.
> We also want to do functional connectivity as I mentioned, which
> requires a beamformer that operates in the time domain as opposed to the
> frequency domain.
> However, there arent any tutorials on the Constrained Minimum Variance
> (LCMV) beamformer (which I think works in the time domain), so I am
> trying to make do with the one used in the tutorial.
> Any help will be greatly appreciated.
> *Tyler Grummett ( BBSc, BSc(Hons I))*
> *PhD Candidate*
> *Brain Signals Laboratory*
> *Flinders University*
> *Rm 5A301*
> *Ext 66124*
> fieldtrip mailing list
> fieldtrip at donders.ru.nl
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the fieldtrip