Back to the main page.
Bug 1323 - should LF column normalisation occur before or after dipole orientation selection?
| Status | ASSIGNED |
| Reported | 2012-02-09 16:17:00 +0100 |
| Modified | 2013-01-23 13:13:39 +0100 |
| Product: | FieldTrip |
| Component: | core |
| Version: | unspecified |
| Hardware: | PC |
| Operating System: | Windows |
| Importance: | P3 normal |
| Assigned to: | Johanna |
| URL: | |
| Tags: | |
| Depends on: | |
| Blocks: | |
| See also: |
Johanna - 2012-02-09 16:17:51 +0100
A brief test of this that I did in Nutmeg about a year ago showed me that there are differences between the two, particularly near the surface of the head but not as much in the centre. If indeed different, which is better? If 'after' is better, then change to code needed in beamformer_lcmv (if projectmom is yes) to do the LF norming within this function after first determining the SVD orientation (and not in ft_prepare_leadfield). I have not yet tested this with FT, but mention this now, as possibly related to bug 1321?
Sarang Dalal - 2012-02-09 16:54:06 +0100
I tested this question a looong time ago, around 2005, so my memory is a bit hazy. But the order of these procedures can indeed make a big difference in the results, and I think Johanna's depth relationship makes sense. I do recall that both with simulations and real data, that the performance was "better" the way it is currently implemented in Nutmeg; i.e., projection of SVD-estimated orientation occurs first, then lead field normalization (if enabled). I believe this procedure would also make it equivalent to using a vector beamformer, and projecting the final results onto the SVD-estimated orientation. It's been way too long for me to have any hope to find my experiments and show you the evidence... but one can also perform a thought experiment: suppose you had a 3-component spherical MEG head model. For the sake of argument and to create the most painful hypothetical circumstances, further suppose that you've rotated your vectors such that the 3rd column contains the radial component... i.e., is exactly zero. Even Matlab will beg you not to normalize your 3-column lead field in this case. ;-) Also in the more forgiving case of a rank-3 BEM lead field, you could still have a large proportion of voxels that have tiny values in one direction, throwing off the normalization. So, in my judgement, projecting onto a best-fit orientation first protects against this horrific outcome... Sarang