Wrong labeling upper cervical cord

Marco10 · February 5, 2025, 9:30am

Dear SCT community,

I am trying to segment and label T1-w scans containing a small portion of the upper cervical cord.
The segmentation is working fine, but the vertebral labeling sometime is failing, as in this case here:

Additionally, in some cases, both the segmentation and the vertebral labelling are producing an output, but the labeling is not always delivering the expected results.On the left a correct labeling, on the right something went wrong.

Is there a way to improve the labeling without manual intervention (since I have thousands of scans to process)?

jcohenadad · February 5, 2025, 3:31pm

Hi Marco, did you try the new TotalSpineSeg method? Available as of v6.5 (see changelog).

Marco10 · February 5, 2025, 3:43pm

Hi, do you mean the contrast agnostic method for segmentation sct_deepseg ?

jcohenadad · February 5, 2025, 3:58pm

No, I meant TotalSpineSeg, which does the vertebral labeling.

Marco10 · February 5, 2025, 4:04pm

Do I have to segment first and then run labeling by running the two command sequentially ?
sct_deepseg -task seg_sc_contrast_agnostic -i $image_file
sct_deepseg -task TotalSpineSeg -i $image_file

jcohenadad · February 5, 2025, 4:08pm

no, you can directly input the image (more details in the doc)

Marco10 · February 5, 2025, 4:17pm

Ok, I will try, thanks.
Anyway, according to this note GitHub - neuropoly/totalspineseg at r20241005, we are not suppose to use this tool for CSA estimation. Is it correct?

Important Note: While TotalSpineSeg provides spinal cord segmentation, it is not intended to replace validated methods for cross-sectional area (CSA) analysis. The spinal cord segmentation from TotalSpineSeg has not been validated for CSA measurements, nor has it been tested on cases involving spinal cord compressions, MS lesions, or other spinal cord abnormalities. For accurate CSA analysis, we strongly recommend using the validated algorithms available in the Spinal Cord Toolbox.

jcohenadad · February 5, 2025, 4:56pm

absolutely correct-- you can use TotalSpineSeg just to get the levels right, and use contrast-agnostic (or deepseg_sc) to get the CSA

Marco10 · February 5, 2025, 5:30pm

TotalSpineSeg doesn’t output the different segments in the cord. The cord is segmented as one mask and the vertebrae C1 and C2 are a unique ROI.

When computing CSA, don’t we need an image with the levels?

sct_process_segmentation -i t2_seg.nii.gz -vert 3:4 -vertfile t2_seg_labeled.nii.gz -o csa_c3c4.csv

Is there a way to extract the different levels from the output of totalspineseg?
sct_deepseg -task totalspineseg -i $image_file -qc qc_singleSubj

Instead of using sct_label_vertebrae?
sct_label_vertebrae -i $image_file -s deepseg/${base_filename:0:-7}_seg.nii.gz -c t1 -qc qc_singleSubj

Nathan_Molinier · February 5, 2025, 7:03pm

Hi Marco, I developed TotalSpineSeg and added the method to SCT.

From what I understand, you would like to use TotalSpineSeg to extract the discs levels. To do that you need to:

Use the output with the suffix _step1_levels generated when running totalspineseg:

sct_deepseg -task totalspineseg -i $image_file # ---> Get the file with **_step1_levels**

Segment the spinal cord
Use the function sct_label_vertebrae with the flag -discfile

sct_label_vertebrae -i $image_file -s -i $sc_seg -c t? -discfile $file_with_step1_levels

Marco10 · February 5, 2025, 7:58pm

Hi Nathan,
Thanks for helping out.

The problem I am experiencing, is that the coverage of the cord is limited to c3/c4 more or less.

Initially, I run a spinal cord segmentation using the: sct_deepseg_sc and than run:

sct_label_vertebrae -i t2.nii.gz -s t2_seg.nii.gz -c t2 -qc ~/qc_singleSubj

The segmentation is working fine. The contrast agnostic approach (which I also tried) is much faster compared to the standard deepseg.

The problem is the sct_label_vertebrae is not working well for images with such small coverage of the cord.

Coming back to your suggestion. Using the output from the totalspineseg model you implanted, would it be possible to have the possibility to split c1 and c2?

Nathan_Molinier · February 5, 2025, 8:17pm

If you do as I suggested, you will get this final output:

Marco10 · February 10, 2025, 1:29pm

Hi Nathan,

I run the code as you suggested.
The segmentation is generated, the totalseg output generated, but when i combine the step1_level with the segmentation I get this error and cannot extract the segmented portions of the cord:

Apply straightening to segmentation...
File /tmp/sct_2025-02-10_08-06-06_label-vertebrae_eczyqo9h/segmentation_straight.nii already exists. Will overwrite it.

Apply straightening to disc labels...
Creating temporary folder (/tmp/sct_2025-02-10_08-06-36_apply-transfo-3d-label_qjogmjkl)
File labeldisc_straight.nii.gz already exists. Will overwrite it.
During disc labeling, center of mass calculation failed due to discontinuities in segmented spinal cord; please check the quality of your segmentation. Using interpolated centerline as a fallback.
Traceback (most recent call last):
  File "/usr/prog/software/SCT-toolbox/6.5/spinalcordtoolbox/scripts/sct_label_vertebrae.py", line 507, in <module>
    main(sys.argv[1:])
  File "/usr/prog/software/SCT-toolbox/6.5/spinalcordtoolbox/scripts/sct_label_vertebrae.py", line 365, in main
    label_vert('segmentation_straight.nii', 'labeldisc_straight.nii.gz')
  File "/usr/prog/software/SCT-toolbox/6.5/spinalcordtoolbox/vertebrae/core.py", line 42, in label_vert
    label_discs(fname_seg, discs)
  File "/usr/prog/software/SCT-toolbox/6.5/spinalcordtoolbox/vertebrae/core.py", line 514, in label_discs
    cx, cy = [int(x) for x in np.round(center_of_mass(slices)).tolist()]
  File "/usr/prog/software/SCT-toolbox/6.5/spinalcordtoolbox/vertebrae/core.py", line 261, in center_of_mass
    raise EmptyArrayError("Center of mass can't be calculated on empty arrays.")
spinalcordtoolbox.types.EmptyArrayError: Center of mass can't be calculated on empty arrays.

Nathan_Molinier · February 10, 2025, 2:33pm

Hi Marco,

what command did you run exactly to generate this error ?

Marco10 · February 10, 2025, 2:40pm

sct_deepseg -task totalspineseg -i $image_file -qc qc_singleSubj

sct_deepseg -task seg_sc_contrast_agnostic -i $image_file -qc qc_singleSubj

The error comes running this:
sct_label_vertebrae -i $image_file -s ${base_filename:0:-7}_seg.nii.gz -c t1 -discfile ${base_filename:0:-7}_step1_levels.nii.gz -qc qc_singleSubj

Nathan_Molinier · February 10, 2025, 3:18pm

The commands look ok, that’s strange.

Could you check if ${base_filename:0:-7}_seg.nii.gz or ${base_filename:0:-7}_step1_levels.nii.gz is empty ? Using your qc_singleSubj

Marco10 · February 10, 2025, 3:37pm

the segmentation is fine and also the step1_levels

jcohenadad · February 10, 2025, 3:46pm

@Marco10 if you send us your data and the exact code you ran, we will be able to help you efficiently

Marco10 · February 20, 2025, 8:10pm

Sorry for replying so late, I had to double check all the files and the image leading to that error was affected by severe ghosting artifacts thus unusable.

In this regard, are there quality checks you would suggest running before implement the segmentation and subsequent CSA estimation ? Is there anything automated in SCT for a dataset of more than 50’000 scans?

In this publication from your group, gradient related non-linearities are corrected for reduced FOV scans.
Is this artifact removal present in SCT at the moment ?

jcohenadad · February 20, 2025, 8:20pm

Is there anything automated in SCT for a dataset of more than 50’000 scans?

if by ‘anything’ you mean ‘quality control’ then the answer is no, we don’t have any automated QC tools. What I suggest is that you look at the QC report for a random selection of subjects, for each site, to get a sense of what the problem could be with some sites.

Is this artifact removal present in SCT at the moment ?

no, it requires proprietary information from the MR vendor. See more info in the paper.