Testing contrast-agnostic model on GRE magnitude data

[jcohenadad]

As of SCT version spinalcordtoolbox/spinalcordtoolbox@bb479d8 (install dev version until SCT v6.4 is released)

Syntax:

# Cord segmentation
sct_deepseg -i sub-test5_magnitude1.nii.gz -task seg_sc_contrast_agnostic -qc qc
# And then to dilate
sct_maths -i sub-test5_magnitude1_seg.nii.gz -dilate 5 -shape disk -dim 1 -o sub-test5_magnitude1_seg_dil.nii.gz

Red: contrast-agnostic (release https://github.com/sct-pipeline/contrast-agnostic-softseg-spinalcord/releases/tag/v2.4), Green: sct_deepseg_sc:

With the mask:

@chaigner

[jcohenadad]

Trying with a lumbar data:

First thing to do is to reorient the image (bc AP and SI are swapped):

sct_image -i 109_Rekos_magnitude1.nii -transpose y,x,z -o 109_Rekos_magnitude1_transposed.nii.gz
sct_image -i 109_Rekos_magnitude1_transposed.nii.gz -flip x -o 109_Rekos_magnitude1_transposed.nii.gz

Which gives:

Now we can run the inference:

sct_deepseg -i 109_Rekos_magnitude1_transposed.nii.gz -task seg_sc_contrast_agnostic -qc qc

Result (room for improvement 😅):

QC report: qc.zip

Note

For creating the mask for shimming, binary segmentation will suffice. However, for precise evaluation of shimming methods, e.g., computing B0 inside the spinal cord, then the soft segmentation should be used (see entry "2024-06-05 10:26:01" in the QC report).

[naga-karthik]

I added a comparison with the prediction from the new model. The gif below shows the comparison. Show in red is model M5prime, and shown in green is nnunet3D

We get a decent segmentation with the nnUNet3D model (green). Note that this is completely out-of-distribution.

[jcohenadad]

#108 (comment)

@naga-karthik can you please replace 'new' and 'old' with the actual names of the model in your post? (eg: M5', nnUnet3D_20250205, etc.) thanks!

[naga-karthik]

Done now!

[jcohenadad]

We get a decent segmentation with the nnUNet3D model (green). Note that this is completely out-of-distribution.

cool! although i would be cautious about interpretation-- there seem to be some oversegmentation at the bottom

[naga-karthik]

there seem to be some oversegmentation at the bottom

indeed ! the nnunet prediction is relatively better as it predicts something rather than an empty pred