Yes, we see the same behavior in verbose mode with small initial negative values leading to much larger deviations after it attempts to correct them into range.
Mike
________________________________________
From: FSL - FMRIB's Software Library [[log in to unmask]] on behalf of Duncan Cleveland [[log in to unmask]]
Sent: Thursday, June 18, 2015 8:44 AM
To: [log in to unmask]
Subject: Re: [FSL] Problems with FNIRT in TBSS_2
Mike,
I was able to test changing the jac range, and this seemed to fix the bad output fine. Thank you for that suggestion.
I can also confirm that the problem appears to be in FNIRT correcting negative jacobian values. I ran TBSS_2 on one of my bad subjects with the verbose output, and found this in the log files:
Jacobian range is -0.0349606 -- 2.0205
Forcing Jacobian range to 0.01 -- 100
Jacobian range is -0.447054 -- 2.08704
Forcing Jacobian range to 0.01 -- 100
Jacobian range is -0.818085 -- 2.1032
Forcing Jacobian range to 0.01 -- 100
Jacobian range is -0.869137 -- 2.47594
Forcing Jacobian range to 0.01 -- 100
Jacobian range is -2.87232 -- 3.27142
Forcing Jacobian range to 0.01 -- 100
Jacobian range is -2.74864 -- 6.32773
Forcing Jacobian range to 0.01 -- 100
Jacobian range is -6.10168 -- 11.5626
Forcing Jacobian range to 0.01 -- 100
Jacobian range is -4.9225 -- 17.3458
Forcing Jacobian range to 0.01 -- 100
Jacobian range is -5.99054 -- 27.8079
Forcing Jacobian range to 0.01 -- 100
Jacobian range is -8.75022 -- 41.8828
Forcing Jacobian range to 0.01 -- 100
Jacobian range is -6.37181 -- 69.1264
Do you find something similar when you run your subjects through FNIRT with the verbose output? And can anyone from FSL speak to the source of this error? It looks like there may be a bug in FNIRT if the process of fixing the Jacobian range leads to even worse range values, which appears to produce poor warp fields.
Duncan
On Thu, Jun 11, 2015 at 12:22 PM, Angstadt, Mike <[log in to unmask]<mailto:[log in to unmask]>> wrote:
Duncan,
Agreed, I don’t actually like the extra interpolation. It was just something I happened upon while testing different things to diagnose the problems.
A colleague and I have been testing, and he has found that adjusting the jacrange to be a bit more tolerant (-1 to 100 instead of the default 0.01 to 100) can fix some subjects as well. In observing what happens during the warping it seems that some subjects have an initially very small negative jacobian value, but then in the process of correcting that and subsequent steps it gets progressively worse. But if we tolerate very small negative values, these don’t get progressively worse and we end up with a warp that appears good, for at least some subjects.
I’m still wondering what the initial problem is, given that simply applying the affine transform (and interpolating) solves it as well.
-Mike
From: FSL - FMRIB's Software Library [mailto:[log in to unmask]<mailto:[log in to unmask]>] On Behalf Of Duncan Cleveland
Sent: Thursday, June 11, 2015 12:13 PM
To: [log in to unmask]<mailto:[log in to unmask]>
Subject: Re: [FSL] Problems with FNIRT in TBSS_2
Hi Mike,
I tried your method of manually applying FNIRT to the FLIRT output, and I was able to successfully have the scans warp to the target. However, this approach does involve an extra interpolation step that ideally should be avoided.
I assumed this extra smoothing fixed some problem in the underlying data, and so I reran TBSS using input FA images that had been smoothed with a 2mm kernel. Unfortunately, that did not produce good warps. I still wonder if there are correctable problems in the input FA images that should be visible to TBSS users, as applying extra interpolation doesn't seem desirable as a general strategy.
-Duncan
On Tue, Jun 9, 2015 at 12:40 PM, Angstadt, Mike <[log in to unmask]<mailto:[log in to unmask]>> wrote:
Hi Duncan,
I just posted about a similar issue yesterday.
In my case, if I manually apply the FLIRT transform to the data, and then perform the same FNIRT that the tbss_2_reg function would have used (you can find the commands in fsl_reg that tbss calls), my warping problems are resolved, so you might check if the same is true for you. I'm not sure why this is the case, as I would think it should yield very similar results in either case since the affine transform from FLIRT is used as an initial transform before the FNIRT step, but something about actually applying it to create a new image before FNIRT fixes things.
Still trying to figure out what's actually going on to cause the bad warps in the first place though, as I similarly don't have any obvious differences between good and bad subjects.
-Mike
> -----Original Message-----
> From: FSL - FMRIB's Software Library [mailto:[log in to unmask]<mailto:[log in to unmask]>] On
> Behalf Of Duncan Cleveland
> Sent: Tuesday, June 09, 2015 8:56 AM
> To: [log in to unmask]<mailto:[log in to unmask]>
> Subject: [FSL] Problems with FNIRT in TBSS_2
>
> Dear FSL experts,
>
> I am currently running diffusion scans through TBSS (12 directions, b=1000
> s/mm2, 2mm isotropic voxels, 64 slices). Most of the output was generated
> correctly. However, there were a few cases (about 5%) where FNIRT
> generated output that was “warped up like a discarded crisp bag” to use a
> description from the FSL email archives (https://www.jiscmail.ac.uk/cgi-
> bin/webadmin?A2=fsl;fa4a69c8.1305). As in the referenced email, error
> messages were produced referring to Jacobian values well outside the
> prescribed range, particularly at the low end where values fell between -4 to
> -9. In all cases the warped output also is stretched far beyond the brainmasks
> of the input and reference images. Unfortunately, we haven’t been able to
> locate obvious problems or artifacts in the raw images, or in the output after
> the FLIRT step (using the FMRIB58_FA_1mm standard template, and the –
> searchy and –searchz options to rotate the raw images). The input FA images
> do have regions of high values running along the perimeter of the brain (due
> to imperfect BET results) as well as truncated coverage of the cerebellum in
> the scan box for some cases. However, most of our raw FA images have
> seemingly identical flaws but made it through TBSS processing without
> problems, and using extra erosions and smaller input brainmasks did not fix
> the problematic cases.
>
> If requested, I can provide cases illustrating successful and unsuccessful TBSS
> processing with similar-looking input FA images, including the input images,
> the images after FLIRT, the warp fields, and the warped images produced by
> the tbss_3 step.
>
> Thanks in advance,
>
> Duncan
**********************************************************
Electronic Mail is not secure, may not be read every day, and should not be used for urgent or sensitive issues
**********************************************************
Electronic Mail is not secure, may not be read every day, and should not be used for urgent or sensitive issues
**********************************************************
Electronic Mail is not secure, may not be read every day, and should not be used for urgent or sensitive issues
|