Jesper's answer was the better one, I did not look closely at what you were
doing. For the downsampling, you can chose what method you use. I'm not
sure what FNIRT does internally when you set the subsampling to something
more than one, but if you downsample the image yourself with FLIRT you can
use trilinear, nearestneighbour, or sinc. I tend to use nearestneighbour
for things like ROIs or segmentation results, but trilinear or sinc for
image files (sinc can be more accurate, but takes a lot longer).
Peace,
Matt.
-----Original Message-----
From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On Behalf
Of John Kuster
Sent: Thursday, April 30, 2009 12:42 PM
To: [log in to unmask]
Subject: Re: [FSL] FNIRT failed
Hi Matt,
Thanks for your response, I didn't crop the image so I can see now how that
would add to the amount of computing needed.
Just wondering, when downsampling an image from 1x1x1 to 2x2x2, conceptually
how does that change the actual data value for a single voxel?
Thanks!!
Jake
On 4/29/09 3:01 PM, "Matt Glasser" <[log in to unmask]> wrote:
> That means you ran out of memory. I would guess you are using a 32bit OS.
> At this point, people in the neuroimaging field really need to be moving
to
> 64 bit OSes, as the 2GB per process (and 4GB total) limits can get in the
> way of a lot of analyses. That being said, there may be things you can do
> to reduce FNIRT's memory usage:
>
> 1) Make sure the resolution of the reference image is not more than
2x2x2mm
> for a human brain. You can do this either by downsampling the reference
> image (like what is done for the standard T1 to MNI nonlinear registration
> config file, which uses the 2mm MNI template) or by specifying a
subsampling
> of 2 in the final iterations of FNIRT, rather than 1 (like what is done in
> the TBSS FA to FA standard config file).
>
> 2) Make sure the FOV of the reference image does not contain a lot of neck
> and empty space (i.e. crop it like the MNI template is cropped).
>
> These steps should increase the speed of your FNIRT processing while
> reducing memory usage, without any adverse effect to registration quality.
>
> Peace,
>
> Matt.
>
> -----Original Message-----
> From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On Behalf
> Of John Kuster
> Sent: Wednesday, April 29, 2009 1:35 PM
> To: [log in to unmask]
> Subject: [FSL] FNIRT failed
>
> Hello FSL experts,
>
> I tried to run the following command,
>
> Command:
> fnirt --ref=swap_CTRL02_T1.nii.gz --in=swap_CTRL02_ASL.nii.gz
> --aff=CTRL02_ASL2T1_auto1.mat --cout=CTRL02_ASL2T1_auto2.mat
>
>
> And got an error. The result is below,
>
> result:
>
> Setting subsampling
> Setting reg mode
> Setting lambda
> Setting subsampling
> Setting reg mode
> Setting lambda
> Error occured during estimation at subsampling level > 1
> Exception thrown with message: St9bad_alloc
>
>
> Could anyone tell me what this error means? In the archives it looks like
> st9bad has to do with memory problems, but I am not sure that is the same
> error as here,
>
> thanks in advance!!!
>
> Jake
>
>
|