Hi,
OK, that makes more sense.
I rarely use the -usesqform option in this way as I find that the coordinates are generally unreliable, especially for getting good alignment. So the much more common case is to do intensity-based alignment. But I can imagine that your situation might be not be typical. However, that's why the documentation and GUI options are lacking on this point. I've now added something into the wiki documentation and will make a change to the FLIRT usage message in a future patch.
All the best,
Mark
On 10 May 2013, at 08:06, "Wood, Tobias" <[log in to unmask]> wrote:
> Dear Mark,
>
> "-usesqform -applyxfm" gives the result I want.
>
> I had actually tried "-usesqform -init $FSLDIR/etc/flirtsch/ident.mat -applyxfm", but this doesn't work and now I understand why. flirt's usage info says "-applyxfm (applies transform (no optimisation) - requires -init)" so I was under the impression that you HAD to give the -init option as well. Perhaps it could be altered to say " requires -init or -usesqform"? I also don't see an option in the ApplyXFM GUI to use the qform instead of the identity matrix, and there's no mention of usesqform on the flirt wiki page.
>
> Many thanks for getting this sorted out, it's going to be very handy indeed.
>
> Toby
>
> On 9 May 2013, at 22:21, Mark Jenkinson wrote:
>
>> Dear Tobias,
>>
>> This doesn't really make sense to me in the context of how FSL works.
>> Let me explain what I mean.
>>
>> To get alignment between images we would do it in one of two ways:
>> 1 - run a registration tool (e.g. FLIRT) to derive the transformation matrix on the basis of the image contents (intensities)
>> 2 - use the nifti coordinates to define the transformation matrix (assuming that you can trust the coordinates to give you good alignment of the anatomy, which is a fairly big assumption)
>>
>> In both cases you would resample an image according to the transformation matrix and then the result would get the same FOV, voxel size, and qform/sform information as the reference image used in the transformation (where the transformation is typically done with flirt, fnirt, or applywarp).
>>
>> For the second case, the transformation matrix is determined using the -usesqform option in flirt (along with -applyxfm).
>>
>> I don't understand why you "need to resample them so the image origins are the same" if you are working in FSL. The origins would be the same after transformation in either of the above cases. Maybe this indicates that you want to use the origin/coordinate information to derive the transformation, which would be the second case above.
>>
>> At any rate, I don't really see what your problem is exactly, but I hope the above might help.
>>
>> All the best,
>> Mark
>>
>>
>>
>>
>> On 9 May 2013, at 18:00, "Wood, Tobias" <[log in to unmask]> wrote:
>>
>>> Hi Mark,
>>> I fear I may have obscured the real issue by mentioning fslroi. I definitely do have to worry about this - the problem occurs even if I process the entire image without any cropping. I included the fslroi commands because it allows you to demonstrate the problem on a single image.
>>>
>>> The fundamental problem is that for some of my subjects, the separate scans are not aligned in data/voxel space because they were acquired with different origins. If I view them in fslview, they appear displaced relative to each other by a few voxels. I need to resample them so the image origins are the same and the images match up in fslview. flirt + an identity transform seems to be the standard way to do this in fsl.
>>>
>>> Your description of the problem with the FSL origin vs NIFTI voxel origin appears to be the perfect explanation. The question now is can I work round this?
>>>
>>> SPM's reslice command does not have this issue, and resamples the issues correctly. However my pipeline is currently fsl based and I really don't want to have to fire up matlab for this step.
>>>
>>> Thanks,
>>> Toby
>>>
>>> On 9 May 2013, at 17:09, Mark Jenkinson wrote:
>>>
>>>> Hi,
>>>>
>>>> I'm not sure what's wrong with using the direct output of fslroi, as FSL doesn't require consistent qform/sform matrices between images. It is possible you just don't need to worry about any of this.
>>>>
>>>> All the best,
>>>> Mark
>>>>
>>>>
>>>> "Wood, Tobias" <[log in to unmask]> wrote:
>>>>
>>>>
>>>> Hi Mark,
>>>> The commands I gave aren't what I'm actually doing - they're just a minimum test case to demonstrate the problem.
>>>>
>>>> I have multiple images per subject that I am processing simultaneously. The real aim is to crop my images to contain just the brain instead of the whole rodent head, for performance/disk space reasons. For most of the subjects fslroi by itself would be fine for this, as all the scans have the same FoV/qform/sform etc.
>>>>
>>>> Unfortunately, for a few of the subjects the FoV was messed up slightly between scans. The only thing that changed is the displacement, not the rotation or scaling. To do voxel-wise processing of those subjects, I need to resample/interpolate them to a consistent FoV - hence the calls to flirt as well as fslroi.
>>>>
>>>> Is there a better way to do this? Or would a minus sign or two added to the identity matrix fool flirt into making the origins co-incident?
>>>>
>>>> Thanks,
>>>> Toby
>>>>
>>>> On 9 May 2013, at 15:23, Mark Jenkinson wrote:
>>>>
>>>>> Hi,
>>>>>
>>>>> It is as I thought - you have neurologically ordered images (a positive determinant for the qform or sform matrices). In this case the internal FSL origin (aligned by FLIRT when you specify an identity matrix) is not coincident with the NIFTI voxel origin, but is at the maximum x-coordinate.
>>>>>
>>>>> I'm not quite sure why you'd want to do those FLIRT calls though.
>>>>> There are probably easier ways to achieve your desired goal that a mixture of fslroi and flirt.
>>>>> Feel free to explain the fundamental problem you are trying to solve.
>>>>>
>>>>> All the best,
>>>>> Mark
>>>>>
>>>>>
>>>>>
>>>>> On 9 May 2013, at 15:04, "Wood, Tobias" <[log in to unmask]>
>>>>> wrote:
>>>>>
>>>>>> Hi Mark,
>>>>>> Please find the headers attached. This is from a rodent subject (hence the small voxel sizes), but I have tested this on a human subject as well and seen the same problem. This particular scan was a sagittal acquisition.
>>>>>> Thanks,
>>>>>> Toby
>>>>>>
>>>>>> On 9 May 2013, at 14:25, Mark Jenkinson wrote:
>>>>>>
>>>>>>> Hi,
>>>>>>>
>>>>>>> This most likely depends on whether your images are radiologically or neurologically stored.
>>>>>>> Can you send us the output of fslhd on the INPUT, REF and OUT images?
>>>>>>>
>>>>>>> All the best,
>>>>>>> Mark
>>>>>>>
>>>>>>>
>>>>>>> On 9 May 2013, at 12:39, "Wood, Tobias" <[log in to unmask]> wrote:
>>>>>>>
>>>>>>>> Dear list,
>>>>>>>>
>>>>>>>> I discovered the following while trying to resample some scans of the same subject, but with slightly different FoVs, to a cropped sub-volume of one of those scans. The following steps should reproduce the problem with any image, e.g. if INPUT.nii is 128x128x128, with FSL 5.0.0
>>>>>>>>
>>>>>>>> fslroi INPUT REF_0 0 64 0 64 0 64
>>>>>>>> fslroi INPUT REF_64 64 64 64 64 64 64
>>>>>>>> flirt -in INPUT -ref REF_0 -out OUT_0 -init $FSLDIR/etc/flirtsch/ident.mat -applyxfm
>>>>>>>> flirt -in INPUT -ref REF_64 -out OUT_64 -init $FSLDIR/etc/flirstsch/iden.mat -applyxfm
>>>>>>>>
>>>>>>>> I would expect OUT_0 and OUT_64 to be perfect copies of REF_0 and REF_64 respectively. However, on my system, OUT_64 is a perfect copy of REF_0. This leads me to believe that the offsets contained within the qform/sform are not being respected properly, even though they all appear to set correctly in the various headers.
>>>>>>>>
>>>>>>>> If I run flirt with the -v option, I get the following warning:
>>>>>>>> WARNING: Both reference and input images have an sform matrix set
>>>>>>>> The output image will use the sform from the reference image
>>>>>>>>
>>>>>>>> Is there anyway I can get OUT_64 to match REF_64? Is this a bug?
>>>>>>>>
>>>>>>>> Thanks in advance.
>>>>>>>>
>>>>>>>> Dr Tobias Wood
>>>>>>>> King's College London, Department of Neuroimaging
>>>>>>>> [log in to unmask]
>>>>>> <input_hd.txt><out0_hd.txt><out96_hd.txt><ref0_hd.txt><ref96_hd.txt>
>>>>>
>>>>
>>>
>>> Dr Tobias Wood
>>> King's College London, Department of Neuroimaging
>>
>
> Dr Tobias Wood
> King's College London, Department of Neuroimaging
|