[Neuroimaging] [Nibabel] Public dataset generates unexpected

[Matthew Brett]

Hi Elef,

On Sun, Sep 10, 2017 at 9:04 PM, Eleftherios Garyfallidis
<elef at indiana.edu> wrote:
> Clearly the title was meant to be "Public dataset generates unexpected
> affine matrix" :)
> Apologies,
> Eleftherios
>
> On Sun, Sep 10, 2017 at 1:31 PM Eleftherios Garyfallidis <elef at indiana.edu>
> wrote:
>>
>> Hello Matthew and all,
>>
>> I downloaded a dataset from NITRC  by Boekel et al.
>> https://www.nitrc.org/projects/dwi_test-retest/
>>
>> and I used nibabel to get the affine and voxel size. The authors claimed
>> that the voxel size is 2x2x2mm^3 however the affine tells a different story.
>>
>> import nibabel as nib
>>
>> img = nib.load('pp26_dwi_run01_A.nii.gz')
>>
>> print(img.affine)
>> [[  -1.999    0.047    0.077  108.292]
>>  [   0.042    1.984   -0.471  -94.461]
>>  [   0.047    0.251    3.703 -116.857]
>>  [   0.       0.       0.       1.   ]]
>>
>> nib.affines.voxel_sizes(img.affine)
>> array([ 2.   ,  2.   ,  3.733])
>>
>> The zoom function gives a different answer in agreement with the authors'
>> claim.
>>
>> img.header.get_zooms()[:3]
>> (2.0, 2.0, 2.0)
>>
>> Could it be that the authors damaged the header during preprocessing?
>> I am assuming here that nibabel is bringing the correct information i.e.
>> whatever is in the Nifti1 image.
>>
>> If you agree that this is an issue with the data itself. It would be nice
>> to contact the authors.

So - it looks like the authors have modified the sform field to set a
new affine matrix.   You can check with e.g. "print(img.header)"
Maybe when they say "voxel size" they mean the values set in the
"pixdim", specifically?   Anyway, if I were the authors, I would want
to know about the discrepancy.  I think you can give them all the
details with the information from "print(img.header)"?

See you,

Matthew