Hi again,

> Thanks for responding. My ref image is a high-SNR 3D that is an average of 6 FSPGRs. My in image has the same dimensions because I already applied the affine transform to it. Perhaps it would be fast to feed in the low-res version of the in image (before the affine transform)?
>
> Additionally, I read your response to Marina, and I'm guessing I have another problem. Both my ref and in images are 3D T1's of the same brain; however, the in image was taken after autopsy and 6-weeks of formalin fixation. The brain has undergone significant structural changes due to the fixation procedure, and the contrasts between white matter and gray matter are significantly different. In fact, after fixation white matter is darker than gray matter. I'm guessing this confuses FNIRT? Perhaps a mutual information objective function would be better, but that doesn't exist at this time, correct? I only see SSE as an objective function in the user guide.

yes, I am afraid that is the case. It sounds like for your application you will need to find some other non-linear registration method that s able to use for example Mutual Information.

Jesper


>
> Running FNIRT with the default settings (i.e. 4,2,1,1) overnight had a total runtime of 8 hours and memory usage of 16Gb. Running at 8,4,2,2 took only 30 mins and 11Gb.
>
> Unfortunately, neither result is useful.
>
> Again, I appreciate your comments.
> Jason
>
> Jason Parker, PhD, DABR
> Director of Diagnostic Physics
> Associate Professor
> Department of Radiology and Imaging Sciences
>
> 950 W. Walnut St., R2 E159B
> Indianapolis, IN 46202-5188
> (317) 274-2072 office
> (941) 737-7549 cell
> (317) 274-1067 fax
> [log in to unmask]
>
>
>
>
> -----Original Message-----
> From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On Behalf Of Jesper Andersson
> Sent: Friday, February 16, 2018 10:38 AM
> To: [log in to unmask]
> Subject: Re: [FSL] FNIRT computation times
>
> Dear Jason,
>
> in general fnirt isn’t really that hungry for neither memory or CPU. Can you please give me an example of how you use it and what the dimensions are of the data (--ref and --in) that you feed into it?
>
> Jesper
>
>> On 15 Feb 2018, at 18:48, Jason Parker <[log in to unmask]> wrote:
>>
>> Hello,
>>
>> I have read the user manual for FNIRT and optimized my calculations accordingly. Even so, I was running into memory constraints on my desktop (which has 12Gb), and so I have moved my code to a supercomputer that can allocate 64Gb of memory for each process. This seems to have solved the memory issues; however, several of the calculations are taking a long time, and I keep having to adjust my wall time and start the calculations all over again - which is wasteful.
>>
>> From what I've been able to gather, FNIRT cannot be run on multiple processors - is that correct?
>>
>> Can anyone think of a way to estimate an FNIRT computation time so that I can set the correct wall time for each registration?
>>
>> Thanks,
>> Jason
>