[pymvpa] searchlight voxel number towards the edge of the brain

J.A. Etzel jetzel at artsci.wustl.edu
Sun Jul 27 05:16:42 UTC 2014


On July 25, 2014 5:35:49 AM Jingwen Jin <jinjingwen111 at gmail.com> wrote:

> 1), When the searchlight sphere get close to the brain edges, it will
> include surely fewer number of voxels. Based on my limited readings, I have
> not seen how people address this issue. Would somebody help explain how
> this should be solved ?
This is one of those details that's not usually mentioned in papers, but 
should be!

With volumetric searchlights the usual practice is to set a minimum number 
of voxels that must be in each searchlight, and not calculate searchlights 
that are smaller. I often use around 10% as a threshold; if a "full" 
searchlight is 30 voxels, I don't allow searchlights smaller than 3 voxels. 
With surface searchlights it's more typical to set searchlight size by 
number of voxels, then adjust radius as needed to obtain the number of 
voxels (e.g. see some of Nick Oosterhof's work).

I don't really like either strategy, but that's what we have now. We know 
there is an interaction between number of voxels and classifier 
(correlation, whatever) behavior, so it makes sense to fix number of 
voxels, but having radically different size (radius-wise) searchlights in 
different parts of the brain is also troublesome for interpretation.


> 2), Most of the MVPA analyses using searchlight approach are using SVM.
> Because of the nature of my hypothesis, I am comparing two correlations:
> e.g., if A-B correlation is higher than A-C. Theoretically, I think it
> makes sense to run correlation based searchlight. But just want to check is
> there a reason for not using correlation based searchlight?
Not that I can think of. You can use any metric in a searchlight that is 
appropriate for your analysis. Many of the issues (sensitivity to 
searchlight radius, coping with spatial variability between people) are the 
same for all metrics.

good luck,
Jo



>
> Any answer will be appreciated !!
>
> Best,
> Frances
>
> --
> Jingwen Frances Jin
> Department of Psychology
> Stony Brook University
>
>
>
> ----------
> _______________________________________________
> Pkg-ExpPsy-PyMVPA mailing list
> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
> http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20140727/f742148f/attachment.html>


More information about the Pkg-ExpPsy-PyMVPA mailing list