Hi Jesper,
Thanks for the answer!
Peace,
Matt.
-----Original Message-----
From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On Behalf
Of Jesper Andersson
Sent: Monday, February 15, 2010 8:52 AM
To: [log in to unmask]
Subject: Re: [FSL] FNIRT, general
Hi Matt,
sorry I missed this one initially.
> I have a question about this iterative template generation process
> (I have
> never tried it myself). Does the final template tend to look like the
> original subject picked, or is it a true average of the anatomical
> patterns?
> When I register a bunch of subjects to MNI152 nonlinear template,
> their
> average ends up looking essentially identical to the template. I am
> wondering how biased this all is to the initialization of the first
> subject
> picked for the registration.
This depends very much on the details how you do it. Let us define one
iteration as registering all subjects to a template (standard brain or
sample mean) once. Let us also assume that the first iteration is run
using some standard template as --ref so as to "jump start" the process.
If on your first iteration you run the registration to a very high
resolution (small ksp and lambda) the sample mean after the first
iteration will look quite similar to the standard brain (e.g. MNI152)
and subsequent iterations will change things relatively little.
If on the other hand you run the first iteration with a relatively low
warp resolution (or even affine) subsequent iterations (to the sample
mean) will have a much larger impact and the final average will be a
better representation of the average of that particular population.
There are no objective rules for what warp resolutions to use for the
different iterations and it is almost more of a craft than a science.
On the other hand it is reassuring that if one starts out with a
population of reasonably young, healthy subjects the final average
will be very similar to the MNI152 regardless of the details of how
one runs the individual iterations. I interpret this as a "regression
to the mean", i.e. given a large enough sample of healthy the average
will look very similar.
Hope that was reasonably clear?
Jesper
|