coherence normalization

[Robert Oostenveld]

On 21 Oct 2004, at 17:23, Tom Holroyd wrote:

> When running a coherence volume using a reference dipole, one
> naturally expects the coherence will be high around the reference
> dipole.
>
> This effect tends to dominate the images.
>
> Is there a way to normalize the coherence volume to eliminate
> this effect?  Perhaps by dividing by the coherence in a "control"
> state?

Hi Tom,

The dominating effect of the refdip is indeed very problematic. I just
happened to have discussed this with Joachim Gross, and I have included
our email exchange below. Please first read that ...

Basically I agree with Joachim, and I don't trust the supdip that is
implemented in FieldTrip's sourceanalysis function. Better test and map
the significance of the difference in coherence between two conditions
using randomization of the trials before the coherence is beamed (that
is implemented in sourceanalysis + sourcestatistics).

Robert

--------------------------------------------------------------------
my question to Joachim was
----------------------------------------------------------------------

Begin forwarded message:

> From: Robert Oostenveld <roberto at smi.auc.dk>
> Date: 1 October 2004 10:26:02 GMT+02:00
> To: Joachim Gross <jgross at uni-duesseldorf.de>
> Subject: dipole suppression
>
> Hi Joachim,
>
> What I always still had to ask you is how you do supression of dipoles
> in DICS, especially in the case of coherence imaging. I have thought
> of two ways of projecting them out:
>
> 1) compute supdip leadfield and its projection on the COV/CSD matrix,
> then project it out of the COV/CSD matrix (which looses 2 or 3 from
> its rank).
>
> 2) compute supdip leadfield and add it to the leadfield of the dipole
> with which is scanned (scandip). Subsequently compute the source
> COV/CSD on those 6 leadfield components and select the 3x3 submatrix
> that corresponds with the scandip to continue the computations with.
>
> Both methods don't really gave me very convincing results. A third
> approach would be to add the supdip leadfield to the (identity) noise
> matrix and project it through the filters. Then nai=pow/noise is
> corrected for the presence of the supdip, but that does not result in
> a supressed source coherence distribution. What is your idea or
> approach for this?
>
> best regards
> Robert
>

----------------------------------------------------------------------
and his answer (Joachim, I hope you don't mind me sharing this on the
list)
----------------------------------------------------------------------

Begin forwarded message:

> From: Joachim Gross <jgross at uni-duesseldorf.de>
> Date: 14 October 2004 17:20:45 GMT+02:00
> To: "robert.oostenveld at fcdonders.kun.nl"
> <robert.oostenveld at fcdonders.kun.nl>
> Subject: dipole suppression
>
> Hi Robert,
>
> sorry for the delay.
>
> The dipole suppression is indeed a complex issue.
> We first implemented it because it facilitates visualization and the
> exact identification of the first
> strongest local maxima.
> Nevertheless, it is quite dangerous because the map is (locally)
> distorted in a non-trivial way.
> We are now trying to move away from suppressing the sources. I think
> it would be better to identify the
> significant local maxima (significance based on
> randomization/permutation).
> But what we are doing at the moment is your approach 3.
> So we add the supdip leadfield to the noise covariance matrix and look
> at pow/noise.
>
> For coherence we are basically doing the same thing.
> So we divide the coherence map (or actually the map of cross spectral
> densities) by a noise map
> that peaks at the locations of the "unwanted" dipoles.
> With this procedure we loose absolute coherence values.
> This is not so important for us since we get the absolute values from
> the coherence and partial coherence spectra
> that are computed afterwards.
> It works surprisingly well but should be used with care.
>
> A better approach would be to map partial coherence (with the unwanted
> dipoles removed). But we have not implemented
> this so far.
>
> Again, I think it is better to have regions of interest identified by
> their significance.
>
> Joachim


----------------------------------------------------------------------
Robert Oostenveld, PhD
Center for Sensory-Motor Interaction (SMI)
Aalborg University, Denmark

and

F.C. Donders Centre for Cognitive Neuroimaging
University Nijmegen
P.O. Box 9101
NL-6500 AH Nijmegen
The Netherlands

Tel: +31 (0)24 3619695
Fax: +31 (0)24 3610989
----------------------------------------------------------------------
N.B. Starting from 1 September 2004, the University of Nijmegen has
changed its name to Radboud University Nijmegen. All web- and
email-addresses ending in ".kun.nl" should therefore be changed into
".ru.nl". Please update your address book and links.