[FieldTrip] megrealign across runs/sessions

[Robert Oostenveld]

Dear Marc,

In principle it is not different to realign the MEG data from three sessions if those sessions were recorded on the same subject (as in your case) or in three different subjects (or a larger number of subjest, as the more common usage scenario). The way you use the function is also correct. However, you should not interpret the residual variance as a measure of the quality of the realignment procedure. The realignment does not know the ground truth, so cannot tell how the attempted realignment matches with the ground truth. It only expresses the difference in the data when the data is compared between original and template position (and when the realignment is repeated but then inversely).

An important aspect however is that you should be careful with applying source reconstruction on the realigned data. In general I would recommend against it, especially if you are planning to apply a beamformer afterwards. The realignment works by making a crude source estimation on on a sheet that is just in the outer grey matter. That crude source estimate (which is a minimum norm estimate with a slight bit of regularization) affects the spatial correlations (i.e. between channels) in the data. Consequently, subsequent source reconstruction might not be able to pick up the spatial correlation any more and might mislocalize. Better (although not always possible) is to source reconstruct the three conditions separate and then combine them.

best
Robert



On 30 May 2012, at 11:47, Marc Recasens wrote:

> Hi everyone,
> 
> I'm considering the possibility to append the MEG data (CTF-275) from 3 different runs (recorded within the same day but with different headpositions in the dewar) into one single dataset. That is, combine my datasets in the sensor-space.
> I've been reading about the possibility to use the ft_megrealign function in order reconstruct the magnetic fields onto a standard gradiometer location,
> However, in the literature this is mainly used to average data across subjects rather than across runs.
> 
> I did a test using the following code:
> cfg= [];
> cfg.template{1} = run1.grad;
> cfg.template{2} = run2.grad;
> cfg.template{3} = run3.grad;
> 
> cfg.vol = vol;   %  single shell headmodel computed from individual MRI
> cfg.inwardshift = 3;
> cfg.verify = 'yes';
> cfg.feedback = 'yes';
> [interp1] = ft_megrealign(cfg, run1);  % trial-based data
> [interp1] = ft_megrealign(cfg, run1);
> [interp1] = ft_megrealign(cfg, run1);
> 
> acording to the results (I show the highest RV), the difference between the original and the realigned data seem really small (which I assume it's good)
> original -> template             RV 2.22 %
> original             -> original RV 2.11 %
> original -> template -> original RV 2.14 %
> 
> 
> I'm wondering whether anyone has experience in using ft_megrealign across runs/sessions and can recommend it (any advise is welcomed).
> According to Knosche (2002), the method seems good but I'd like to know whether someone has test it in real-life situations (especially taking into account the head position differences in the z axis)
> 
> Can affect the accuracy of the subsequent source reconstruction?
> 
> 
> -- 
> Marc Recasens
> PhD Student
> Universitat de Barcelona
> Tel.: +34 639 24 15 98
> 
> _______________________________________________
> fieldtrip mailing list
> fieldtrip at science.ru.nl
> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip