Regenerate the lumbar and top cervical GM mask

[jcohenadad]

possible avenues

• reusing existing chunk, and manually fixing it for better continuity
  • possible issue: the existing chunks are misaligned
• for the top part:
  • select the upper most slice (i) where the GM mask appears,
  • for that slice, register the t2 image to the top slice n,
    • try with the t2 image or the cord mask
  • generate intermediate warps for each slices between slice i and slice n
  • apply warping fields to the GM mask of slice i
• do only one half, and symmetrize later one
  • not needed for previous method

VC: provide GM mask at top cervical level

[jcohenadad]

Given that the lowest slices of the registered T2*w image is not perfect (see below), I suggest selecting not the lowest slice, but rather a slice above, e.g.: z=150

[jcohenadad]

For 2D registration, consider starting from:

PAM50/scripts/create_atlas/create_atlas.m

Lines 305 to 310 in 61dfb41

	% estimate diffeomorphic transformation
	cmd =['isct_antsRegistration --dimensionality 2 --initial-moving-transform ' affine_atlas ' ',...
	'--transform SyN[0.1,3,0] --metric MeanSquares[' templateci_slice_ref_thresh ext ',' mask_nohd ext ',1,4] ',...
	'--convergence 100x10 --shrink-factors 4x1 --smoothing-sigmas 0x0mm ',...
	'--output [' prefix_ants ',' mask_nohd '_affine_warp' ext '] --collapse-output-transforms 1'];
	disp(cmd); [status,result] = unix(cmd); if(status), error(result); end, disp(result)

[jcohenadad]

Also look at what was done in the past to get the top/bottom slices:

https://github.com/spinalcordtoolbox/PAM50/blob/9cfd3dcfa2eb091aefb407a045789a6b67e109d0/scripts/create_atlas/register_AMU_to_PAM.py

[jcohenadad]

Running into the following issue when creating 2D NIFTI files:

Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/Library/Frameworks/Python.framework/Versions/3.13/lib/python3.13/site-packages/nibabel/nifti1.py", line 2006, in __init__
    super().__init__(dataobj, affine, header, extra, file_map, dtype)
    ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Library/Frameworks/Python.framework/Versions/3.13/lib/python3.13/site-packages/nibabel/analyze.py", line 912, in __init__
    super().__init__(dataobj, affine, header, extra, file_map)
    ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Library/Frameworks/Python.framework/Versions/3.13/lib/python3.13/site-packages/nibabel/spatialimages.py", line 515, in __init__
    raise ValueError('Affine should be shape 4,4')
ValueError: Affine should be shape 4,4

Apparently, NIfTI can only be 3D or 4D.

[jcohenadad]

Looks like in the past we tricked NIfTI to be 2D, although I don't find the code (only the comment):
https://github.com/spinalcordtoolbox/spinalcordtoolbox-dev/blob/e49a827533d244e5172f244c0768be830079ccda/dev/atlas/create_atlas/create_atlas.m#L165

Or maybe slice extraction with c3d software actually creates a 2D NIfTI?
https://github.com/spinalcordtoolbox/spinalcordtoolbox-dev/blob/e49a827533d244e5172f244c0768be830079ccda/dev/atlas/create_atlas/create_atlas.m#L148

[jcohenadad]

OK-- turns out the issue was just in the output header-- I simply have to apply sct_image -copy-header to fix that

[jcohenadad]

Here is the preliminary result of the extension of the T2*w template towards the lumbo-sacral cord:

Screen.Recording.2025-09-09.at.12.12.03.mov

The discontinuity aspect could be improved, but I'm afraid the representation of the cord GM will be far from realistic.

An alternative approach, discussed with @vcallot and @valosekj, would be to concatenate another GM template to the AMU15 one. But there we will face other issues, including intensity discontinuity. But it might be worth trying.

[jcohenadad]

An alternative approach, discussed with @vcallot and @valosekj, would be to concatenate another GM template to the AMU15 one. But there we will face other issues, including intensity discontinuity. But it might be worth trying.

To follow-up on this, I'm adding a couple more thoughts.

• If we really want to concatenate the AMU15 with another, lumbo-sacral, template, then we should re-use the same pipeline that was used to generate the original AMU15 template, otherwise the generated GM/WM lumbo-sacral masks would likely have different probability distribution values (see also Improve registration of AMU WM/GM mask and PAM50 template #41 (comment)), which will impact the new version of the WM atlas, which relies on the GM/WM mask.
• Given the amount of work, I'm leaning towards not extending the T2*w image, but only the WM/GM mask, using linear scaling (as shown in Regenerate the lumbar and top cervical GM mask #45 (comment), but it would only be applied to the GM/WM mask).

[valosekj]

Thanks for the insights!

From the discussions I’ve had about the lumbo-sacral template, it seems useful as a first step to get the GM LSE label in the PAM50 space (since GM LSE shows lower inter-subject variability than SC LSE; ref). This label could then be used for PAM50 normalization together with the tip of the cord. I was imagining something like this: register individual lumbo-sacral scans to the PAM50 to generate a lumbo-sacral T2*w template image, segment GM (using Nilser’s model or manually) on this template image, compute slice-wise GM CSA, and identify the GM LSE as the slice with the largest GM CSA. In other words, having a GM mask would already be helpful even without the WM atlas. (Hope I understood the discussions correctly and haven’t missed any crucial details.)

[jcohenadad]

From the discussions I’ve had about the lumbo-sacral template, it seems useful as a first step to get the GM LSE label in the PAM50 space (since GM LSE shows lower inter-subject variability than SC LSE; ref). This label could then be used for PAM50 normalization together with the tip of the cord. I was imagining something like this: register individual lumbo-sacral scans to the PAM50 to generate a lumbo-sacral T2*w template image, segment GM (using Nilser’s model or manually) on this template image, compute slice-wise GM CSA, and identify the GM LSE as the slice with the largest GM CSA. In other words, having a GM mask would already be helpful even without the WM atlas. (Hope I understood the discussions correctly and haven’t missed any crucial details.)

I'm not sure I fully understand the intend here. Are you suggesting to identify a single-point label corresponding to the LSE? And if so, would that label be added to the official list of SCT labels? With value 61 for example?

Or are you suggesting to generate a GM mask with better accuracy than what is proposed in #45 (comment)? If that's the case then yes, I think it would be a good idea to generate a more truthful representation of the GM. In this case, we should generate GM mask for each individual subjects, to mimic what was done on the AMU15 case (in order to build a probabilistic lumbo-sacral GM mask, otherwise it will be binary and that won't be compatible with the AMU15 mask). Also: I was under the impression that there was already a GM mask for these subjects?

having a GM mask would already be helpful even without the WM atlas.

They go hand-in-hand. Once we have the GM mask, we will re-run the WM atlas pipeline to update it, according to the new GM mask (we have to, otherwise atlas/ and template/ files won't match, which would be a problem).

[valosekj]

Are you suggesting to identify a single-point label corresponding to the LSE?

Yes, exactly. The idea is to identify a single label corresponding to the GM LSE based on the GM CSA (i.e., the location will likely differ from the LSE based on the spinal cord CSA).

And if so, would that label be added to the official list of SCT labels? With value 61 for example?

Ideally yes. So that it would also be useful to other users.

Or are you suggesting to generate a GM mask with better accuracy than what is proposed in #45 (comment)? If that's the case then yes, I think it would be a good idea to generate a more truthful representation of the GM. In this case, we should generate GM mask for each individual subjects, to mimic what was done on the AMU15 case (in order to build a probabilistic lumbo-sacral GM mask, otherwise it will be binary and that won't be compatible with the AMU15 mask). Also: I was under the impression that there was already a GM mask for these subjects?

As far as I know, only spinal cord segmentations are available in the native space. If I remember correctly, GM segmentations exist only for some subjects and/or only for part of the FOV. That’s why I suggested segmenting GM directly in the PAM50 template image. However, doing so would indeed produce only a binary mask, which is definitely a limitation. We can discuss how feasible (and for how many subjects) it would be to obtain GM masks in the native space.