search

MRtrix3 / mrtrix3

MRtrix3 provides a set of tools to perform various advanced diffusion MRI analyses, including constrained spherical deconvolution (CSD), probabilistic tractography, track-density imaging, and apparent fibre density
http://www.mrtrix.org
Mozilla Public License 2.0
294 stars 181 forks source link

-export_grad_fsl fix for highly non-RAS acquisitions #2912

Closed Lestropie closed 2 months ago

Lestropie commented 6 months ago

Here we go again...

So I was left scratching my head at this one for quite some time. I am trying to get to #2477 among other things, but was seemingly having issues at the bvec write stage, which meant I couldn't yet trust anything I was observing at the bvec read stage.

  1. Eigen function .rotation() was being used, which is not the same as .linear(). For a DICOM DWI series acquired on a Siemens magnet with a sagittal plane, these two functions yield very different results. I'm sure we've had at least one issue in the past with conflating these two, though haven't had success in finding anything on GitHub (@MRtrix3/mrtrix3-devs anyone with better recollection?).

  2. Code was doing a post-multiply with the image2scanner transform in order to do a scanner2image transformation., rather than constructing the scanner2image transform and using that.

  3. Code was doing a scanner2image transformation based on the transformation for the internal in-memory representation of the header, and then performing subsequent permutations & flips based on the secondary output of the File::NIfTI::adjust_transform() function and the internal strides, knowing that at NIfTI export time there would be similar modifications applied. However the File::NIfTI::adjust_transform() function returns what will become the header transformation of the output image. So it seems to me that if one wants to perform an image2scanner transformation of gradient directions that will correspond to the output image, and one has what will be the transformation of the output image, might as well do it in one step?

I'm reasonably confident given the test data & tool I have:

In the absence of this change, the specific combination of sagittally-acquired DICOM DWI data, running mrconvert to NIfTI / bvec / bval with internal header transform realignment disabled, produces an incorrect result.

@jdtournier I'll fill in more details in a separate email chain; content to leave as a draft until I can explain the broader project there.

Lestropie commented 6 months ago

Realised that the testing I'm doing currently has commented out the evaluation of manually requested output strides. Will leave as draft until my tests have been expanded in scope.

Lestropie commented 2 months ago

Superseded by #3011.