Lumbar Segmentation issues (Deepseg)

SCT
1

Hi everyone,
I’m trying to apply segmentation algorithms to MRI lumbar images.
I’ve tried different combinations of commands and parameters, and the combination that seems to work best is the following:

sct_image -i filename -setorient LPI -o t2LPI.nii.gz
sct_deepseg_sc -i t2LPI.nii.gz -c t2 -kernel 3d -centerline cnn -brain 0 -o outputFilename.nii
sct_deepseg_sc -i t2LPI.nii.gz -c t2 -centerline file -file_centerline outputFilename.nii -brain 0 -o outputFilename2.nii
sct_create_mask -i t2LPI.nii.gz -p outputFilename2.nii -o mask.nii
sct_deepseg_sc -i t2LPI.nii.gz -c t2 -centerline file -file_centerline mask.nii -brain 0 -o outputFilename3.nii
sct_deepseg_sc -i t2LPI.nii.gz -c t2 -centerline file -file_centerline outputFilename3.nii -brain 0 -o outputFilename4.nii

However, I am still getting poor segmentation results. Some non-spinal cord areas are being segmented, while the segmentation of the spinal cord itself is fragmented.

Screenshot from 2024-09-04 11-33-01
Screenshot from 2024-09-04 11-33-01705×903 303 KB

Indeed, I noticed that this approach only works on some images, while errors are raised for others (different errors depending on the image, such as the centerline not being detected using the CNN method). These images are NIfTI images that have been augmented on the z-axis using a Soup-GAN.

thank you for your help!

2

Hi @Alessia_Sabia,

Thank you for reaching out. Have you tried our new contrast-agnostic method, which is more robust than sct_deepseg_sc:

sct_deepseg -i IMAGE -task seg_sc_contrast_agnostic

If that doesn’t work, we also have a dedicated method for lumbar spinal cord segmentation on T2w data (introduced in v6.4)?

sct_deepseg -i IMAGE -task seg_lumbar_sc_t2w
3

Hi!
Thank you for answering.
Yes, I’ve tried all the available algorithms.
In particular, this one, only segment a tiny part of the spinal cord.

4

If you can send me 1-2 representative images I can have a look

5

I need an email address because they are too big to be attached here.
Thank you for your help

6

jcohen@polymtl.ca

7

Hi @Alessia_Sabia,

I confirm that neither sct_deepseg_sc, nor sct_deepseg -task seg_sc_contrast_agnostic nor sct_deepseg -task seg_lumbar_sc_t2w work properly for your data.

I’ve tried a model we are currently working on, called TotalSpineSeg, and it gave satisfactory results:

ezgif-5-074d363cc7

However, the method is not integrated into SCT yet. It will be at some point. In the meantime, you can try it out yourself via the instructions from the repository.

Cheers,
Julien

8

Thank you for your help!