Results on a nnUNet traind over 86 selected cases

[abelsalm]

Here is the last training I did, same procedure as in the last one.

The idea is that I had some oversegmentation on the most inferior part of whole-spine images, so I added 6 images from this dataset with GT. Now it's better. If you want to look at it, the main issue is on the R-L axis where the subarticular foramens are also white and the model often struggles to only segment until the dural sac.

Since now I'm pretty happy with the results, I think it would still be good to add some GT for more robustness.

For now the images come from :

• 12 from sci-paris
• 7 from whole-spine
• 28 from dcm-brno
• 29 from spine-generic
• 10 from dcm-oklahoma

And there is almost as much healthy subject as ill ones.

What would you advice ? @jcohenadad @sandrinebedard @valosekj
If the new generated groundtruth don't require much correction I could try to balance to reach something like 20/30 images from each dataset. I don't know also if there is some kind of minimal training size to ensure that the model is robust enough.

Here is a link to the qc (it's big i tested 300 images !)

[jcohenadad]

Cool! After a quick review, I noticed that most subjects are missing the segmentation for the upper part of the cervical cord, examples here:

FILES_SEG:
    - 'C:\Users\abels\OneDrive\Documents\NeuroPoly\canal_seg\segmentation\training\tenth_training\imagesTs/sub-046_T2w_000_0000.nii.gz'
    - 'C:\Users\abels\OneDrive\Documents\NeuroPoly\canal_seg\segmentation\training\tenth_training\imagesTs/sub-050_T2w_000_0000.nii.gz'
    - 'C:\Users\abels\OneDrive\Documents\NeuroPoly\canal_seg\segmentation\training\tenth_training\imagesTs/sub-2779B4786B_ses-4786B_T2w_000_0000.nii.gz'
    - 'C:\Users\abels\OneDrive\Documents\NeuroPoly\canal_seg\segmentation\training\tenth_training\imagesTs/sub-2804B4632B_ses-2804B_T2w_000_0000.nii.gz'
    - 'C:\Users\abels\OneDrive\Documents\NeuroPoly\canal_seg\segmentation\training\tenth_training\imagesTs/sub-2804B4632B_ses-4632B_T2w_000_0000.nii.gz'
    - 'C:\Users\abels\OneDrive\Documents\NeuroPoly\canal_seg\segmentation\training\tenth_training\imagesTs/sub-2825B4881B_ses-2825B_T2w_000_0000.nii.gz'
    - 'C:\Users\abels\OneDrive\Documents\NeuroPoly\canal_seg\segmentation\training\tenth_training\imagesTs/sub-2825B4881B_ses-4881B_T2w_000_0000.nii.gz'
    - 'C:\Users\abels\OneDrive\Documents\NeuroPoly\canal_seg\segmentation\training\tenth_training\imagesTs/sub-2847B6453B_ses-2847B_T2w_000_0000.nii.gz'
    - 'C:\Users\abels\OneDrive\Documents\NeuroPoly\canal_seg\segmentation\training\tenth_training\imagesTs/sub-3758B6378B_ses-3758B_T2w_000_0000.nii.gz'

I'm curious how this segmentation compares to that from @NathanMolinier 's model-- it would be important to make a comparison. You want to get started with your ground truths using the best available method.

[NathanMolinier]

I agree that it would make sense to compare with totalspineseg. Let me know if you need help @abelsalm