Back to the main page.
Bug 3089 - implement dipole fitting for combined magnetometer and planar gradiometer data
| Status | ASSIGNED |
| Reported | 2016-03-07 11:09:00 +0100 |
| Modified | 2016-03-08 15:44:19 +0100 |
| Product: | FieldTrip |
| Component: | inverse |
| Version: | unspecified |
| Hardware: | PC |
| Operating System: | Mac OS |
| Importance: | P5 normal |
| Assigned to: | Robert Oostenveld |
| URL: | |
| Tags: | |
| Depends on: | |
| Blocks: | |
| See also: |
Robert Oostenveld - 2016-03-07 11:09:49 +0100
as discussed with Lau, it should be based on https://en.wikipedia.org/wiki/Whitening_transformation implemented at the level of the ft_dipolefitting code (since both for scanning and for the nonlinear part).
Robert Oostenveld - 2016-03-07 11:13:28 +0100
And here are some details about the alternative suggested by Luis (CC): ---- On 01 Mar 2016, at 20:34, Luis Garcia Dominguez wrote: I am working on dipole mapping of epileptic spikes using an Elekta Neuromag 306-Channel Meg system. The ft_dipolefiting function works great only when gradiometers or magnetometers are specified but not when both are pooled together. See the file “uncorrected.txt” attached. This file shows that the residual variance for pooled channels is very high and also the distance between the this condition and the MEG only or GRAD only is relatively high, between 1 to 3 cm, when the distance between MEG and GRAD is about 0.3 cm. I also added the correlation between the dipole orientations to show how similar they are. Inspecting the code in “dipole_fit.m” I found out that the problem was due to the large difference in variances between magnetometers and gradiometers, which is not compensated by the computed leadfield. In order to solve this problem I introduced an additional parameter that is passed during the optimization procedure. That is, the function “dipole_fit” passes an extra number to “dipfit_error” using the “param” variable (which by default contain only the dipole position). This extra number is used as a coefficient to scale the magnetometers, when both magnetometers and gradiometers are present. This parameter is also optimized along with the dipole position adjusting the relative amplitude of mag and grads to best fit the data. After introducing this correction the result seems pretty accurate for the pooled condition and in agreement with the MAG and GRAD. See the “corrected.txt” file. I think that this solution of mine is probably not that elegant since the actual coefficient should be independent of the subject but, on the other hand it is more flexible.
Robert Oostenveld - 2016-03-07 11:15:14 +0100
The best matching example dataset that can be used for testing is the one described here http://www.fieldtriptoolbox.org/tutorial/natmeg/dipolefitting#fit_a_dipole_model_to_the_meg_data It could be used to test magnetometer+planar, but also meg+eeg.
Robert Oostenveld - 2016-03-08 14:47:00 +0100
(In reply to Robert Oostenveld from comment #2) I am visiting Stockholm at the moment, and worked on it with Lau this morning. Using the MEG-audodd dataset we made a test script and added the sphering to ft_dipolefitting. It works both with the gridsearch and the nonlinear part. See https://github.com/fieldtrip/fieldtrip/pull/118 It is to be considered whether it should be made available in a more user-friendly manner (e.g. cfg.prewhiten=yes/no) also considering channel-selection. I will discuss this with Jan-Mathijs (CC) when back in Nijmegen, since he implemented very similar stuff in minimumnormestimate. The combination MEGMAG+MEGPLANAR now work, but MEG+EEG not yet. That is only due to the handling of grad and elec structures, which are now mutually exclusive. The test script can easily be extended to also deal with this case, as the EEG dipole fit is also quite nice in this data. @Luis, could you review the test_bug3089 script? You will see the cfg.accuracy option in ft_preprocessing, which ensures that the units of the forward model are consistent with the units of the data. Also, all computations must be done in SI units (meter, Tesla, Volt).