Dear Carlos,
I'm surprised to hear what you are saying about FNIRT.
Did you make sure that both the brain extracted *and*
original images (including the non-brain structures)
were available? It is true that FNIRT does not work
as well when only given the brain extracted images.
And I assume you were using it via FEAT - as this is
probably the best way to begin with.
As for FIRST - the registration that it does is purely
to initialise its search for the subcortical structures.
It does not refine this registration based on the
final subcortical segmentation. So typically the
initialisation is OK, but far from perfect. I am
really surprised that FNIRT would not be doing
better unless your images have poor SNR.
Finally, the timing issue of going from 2mm to 1mm
would be at *least* 8 times longer as there are
8 times as many voxels (each 2x2x2mm voxel
contains 8 separate 1x1x1mm voxels). Plus the
memory then is 8 times larger and this can easily
start causing swapping, and when this occurs the
speed often goes down by one or more orders
of magnitude! So I'm really not that surprised that
it takes a *lot* longer unless you have a very high
RAM machine (in which case it would probably
only be about 10 times slower!)
All the best,
Mark
On 20 Aug 2010, at 21:35, Carlos Faraco wrote:
> Mark,
>
> I think I slightly misread your reply.
>
> Are you saying that using FNIRT would give me better hippocampal
> alignment than using the FIRST FLIRT registrations that are
> optimized for subcortical alignment?
>
> I have previously tested FNIRT as implemented in FEAT for a whole
> brain analysis to see how well the registration worked. I did not
> like the distortion it created it created in the images. Seemed that
> FLIRT was giving better results.
>
> -Carlos
>
|