Problem reorienting DTI images

[mikearcaro]

Hi,

I have some diffusion data that I preprocessed with MRtrix3 then aligned to a template using ANTs. Prior to registration, the RGB maps look correct (i.e., in native space). After transformation using ANTs' recommended steps here, the colors are incorrect. e.g., the corpus callosum is green instead of red. There is little difference between the template-warped RGB maps pre- and post- ReorientTensor (see attached images), so it seems like the reorienting isn't doing what I want it to do. Below, I copy the commands I used as well as screenshots showing the RGB maps pre and post warping. Any advice on how to debug would be much appreciated!

In case relevant, these data were collected on a Bruker 11.7T scanner.

Best,
Mike

Convert ordering of tensor data from MRtrix to ANTs format:

ImageMath 4 dtiComp.nii.gz TimeSeriesDisassemble tensor.nii
i=0
for index in xx yy zz xy xz yz; do
mv dtiComp100${i}.nii.gz dtiComp_${index}.nii.gz
i=$((i+1))
done
ImageMath 3 dtAnts.nii.gz ComponentTo3DTensor dtiComp_

Calculate FA, MD, and RGB maps

ImageMath 3 dtFA_native.nii.gz TensorFA dtAnts.nii.gz
ImageMath 3 dtMD_native.nii.gz TensorMeanDiffusion dtAnts.nii.gz
ImageMath 3 dtRGB_native.nii.gz TensorColor dtAnts.nii.gz

Register to template

antsRegistrationSyN.sh -f template.nii.gz -m bzero.nii.gz -d 3 -o dwi2template -n 4

Warp tensor to template

antsApplyTransforms -d 3 -e 2 -i dtAnts.nii.gz -o tensor2template.nii.gz -t dwi2template1Warp.nii.gz -t dwi2template0GenericAffine.mat -r template.nii.gz

Correct orientation due to warping

antsApplyTransforms -d 3 -o [dtCombinedWarp.nii.gz,1] -t dwi2template1Warp.nii.gz -t dwi2template0GenericAffine.mat -r template.nii.gz

ReorientTensorImage 3 tensor2template.nii.gz tensor2template_reoriented.nii.gz dtCombinedWarp.nii.gz

ImageMath 3 dtFA_template_reorient.nii.gz TensorFA tensor2template_reoriented.nii.gz
ImageMath 3 dtMD_template_reorient.nii.gz TensorMeanDiffusion tensor2template_reoriented.nii.gz
ImageMath 3 dtRGB_template_reorient.nii.gz TensorColor tensor2template_reoriented.nii.gz

Native space:

After antsApplyTransforms but before ReorientTensorImage:

After ReorientTensorImage:

[cookpa]

Sorry I have not had time to look at this. Can you share the template an an example dtAnts.nii.gz?

[mikearcaro]

Hi Philip,

Thanks so much for following up on my post! The files are quite large (>300MB each). I uploaded to a Dropbox folder that can be shared. Should I post the link here of or better to email you?

Best,
Mike

p.s. We're neighbors. I'm at Penn too. 4th floor Goddard. Happy to swing by if easier to chat in person.

[cookpa]

Hi, I think I'm missing template.nii.gz. I have under 2_template the files dtFA_template.nii.gz dtMD_template.nii.gz dtRGB_template.nii.gz template_avg7shrews.nii.gz tensor2template.nii.gz.

Could you please add template.nii.gz, the one used as the fixed image in the registration and reference image in antsApplyTransforms?

Thanks

[mikearcaro]

Thanks for looking at the files! The template_avg7shrews.nii.gz is the template. I just shortened it when posting the code to cut down on filename length. Sorry, that was not clear. I renamed the files on dropbox.

[cookpa]

Got it. I realize I also don't have bzero.nii.gz.

Also, what is your ANTs version? You can tell this with

antsRegistration --version

[mikearcaro]

Added the bzero in native space to the 1 folder. Running version 2.3.5.post130-gaacdc61 on Intel iMac w/ OSX Catalina

[cookpa]

The bzero is a mif file, can you please add the ".nii.gz"

I'm trying to reproduce what went on. I see a mixture of qforms and sforms in the data. There have been changes to the logic of how ITK handles this, so the latest ANTs may work correctly, or at least differently.

[mikearcaro]

Sorry again! I added the nii.gz file to folder 1. Do you think I'm doing something weird in the preprocessing of the diffusion data? Or could there be something weird with how I have the template oriented? The RGB map looks correct in native space after converting from mrtrix to the ANTs format. It gets weird (AP and RL seem to be flipped) when warping to template. Maybe that the template is in a weird space and, as you said, ANTs is warping correctly?

[cookpa]

Thanks for all these files, I think I've ruled out any problems with the data at the tensor level.

I believe the issue is that your input tensors are in physical space. This is why the RGB images in native space look correct - because ITK-SNAP displays the image in it's best guess of the physical canonical orientations - it uses the image header to figure out how to display the image so that the L, R, P, A, I, S, labels are correct.

There's disagreement between software on how to represent the gradient directions (bvecs) and tensors. It's a long story but ANTs expects tensors in voxel space on disk. When you call antsApplyTransforms, it converts them from the local space of the moving image to that of the reference image. ReorientTensorImage should convert them to physical space internally, and apply the reorientation, and output the corrected images in voxel space again.

You can call RebaseTensorImage to rotate the tensors to voxel (LOCAL as it calls it) space. This should make the registration workflow run as expected. After warping and reorienting, you can call RebaseTensorImage again to take them back to physical space, which should make the RGB look correct again in the template space.

I'm sorry to have to warn about a longstanding issue with tensor orientation and reorientation (#1137). I've tried to address this but have not had the resources to do so definitively. You might need to verify the reorientation yourself, by doing a rigid registration that will make problems obvious.

If you load your template into SNAP as an overlay, you can apply a large (eg, 30 degree) rotation - save that and use as a reference image. Then do a registration to the rotated template (using rebased tensors as above), apply reorientation, and then rebase and look at tensor color. This should highlight any problems.

This is all very confusing, so just to reiterate:

• TensorColor doesn't do any transformation of the DT, it just computes the eigensystem and outputs the vector it finds. For registration to work correctly, these colors should agree with the image space, so if your image has sagittal slices, then the midsagittal corpus callosum would be blue because blue is z in voxel space.

• antsApplyTransforms does transform the DT, because it has to resample it into the new image space. It expects input in the voxel space of the moving image, and will rotate the tensors into the voxel space of the fixed image.

• antsApplyTransforms doesn't reorient based on deformation - it just transforms coordinates from the moving image space to fixed image space. In ITK-SNAP, your dtANTs image is ASL, and the template is listed as "Oblique (closest to IPR)". So it wants to change the tensor coordinates between those two coordinate systems. This is independent of any registration, it's just reslicing. I hope this makes sense.

• ReorientTensorImage does care about the registration. It tries to estimate the rotation component of the local warp, and use that to reorient the tensors.

To simplify, imagine we had a consistent voxel space between the moving and template images, like RPI. Then if you called antsApplyTransforms -r template -i input -t Identity, it should not rotate anything and the colors would not change. But if you did a registration, ReorientTensorImage would still have work to do. After ReorientTensorImage, the colors should be different, but the difference should depend on what rotation happened in registration (ie, in the physical space).

Unfortunately, there's also degeneracy in the RGB representation, so it's not easy especially if the rotations are fairly small to judge by eye. The best way to tell is via tractography, but that again requires importing to some other software, which might interpret the image differently.

[cookpa]

FYI simple registration, using only translation of the moving image, seems to work as expected.

RebaseTensorImage 3 dtAnts.nii.gz dtAntsLocal.nii.gz LOCAL  

ImageMath 3 native_fa.nii.gz TensorFA dtAntsLocal.nii.gz 

antsRegistration -d 3 -r [ template_avg7shrews.nii.gz , native_fa.nii.gz , 1 ] -t Translation[ 0.05 ] -m Mattes[ template_avg7shrews.nii.gz , native_fa.nii.gz , 1, 32 ] -f 6x4x1 -s 3x2x0vox -c 5x5x0 -o [ faToTemplateTranslation, faToTemplateTranslationDeformed.nii.gz ] --verbose

antsApplyTransforms -d 3 -e 2 -r template_avg7shrews.nii.gz -i dtAntsLocal.nii.gz -t faToTemplateTranslation0GenericAffine.mat -o dtAntsLocalToTemplate.nii.gz --verbose

RebaseTensorImage 3 dtAntsLocalToTemplate.nii.gz dtAntsPhysicalToTemplate.nii.gz PHYSICAL

ImageMath 3 rgbTemplatePhysical.nii.gz TensorColor dtAntsPhysicalToTemplate.nii.gz

If it's OK with you, I'll hang on to this data for additional testing - I won't share it further without permission, of course.

[mikearcaro]

Hi Philip,

Thanks so much for spending all this time looking at my data and for thoroughly explaining the issues with tensor reorientation. This is all really terrific and super helpful! Running RebaseTensorImage before and after antsApplyTransforms results in reasonable looking RGB maps in template space. I still need to test with the ~30 degree rotation and also see if tracking looks OK.

I also will probably just recreate the template in ASL. There's no reason it needs to be IPR (and not sure why it's oblique in the first place).

Happy for you to hold onto the data and use it for additional testing. Also, happy to provide more data of that kind if it would be useful for testing!

Best,
Mike