Correction of the segmentation on Mild Compression cases

[abelsalm]

Since the model has been mostly trained on healthy subjects, it is not reliable on subjects with compressions of the canal. Those compressions often create concavities that the model hardly detects. So I would like to take some of those outputs and correct them to train the model again on a dataset containing both healthy and unhealthy subjects.

My issue is that in some cases I don't really know where to segment precisely :

Here you have twice the same axial visualisation of sub-hamburg03 preprocessed for nnU-Net, and I can't say on the blue arrows if the segmentation should or not match all the details. Same for the global concavity on the anterior side of the canal.

An other case where CSF disappears on the anterior side, here also I don't know how to deal with this side.

And a last one, where it's more about the global canal shape.

If someone could give me hints on how to deals with those 3 examples, it would be a good model for the rest of the subjetcs. @valosekj @sandrinebedard

[sandrinebedard]

Here you have twice the same axial visualisation of sub-hamburg03 preprocessed for nnU-Net, and I can't say on the blue arrows if the segmentation should or not match all the details. Same for the global concavity on the anterior side of the canal.

here I see the spinal rootlets (in black) at the edge of the spinal canal, I think I would include them, since we want the entire canal cavity, it is normal that it includes the spinal nerve rootlets.

An other case where CSF disappears on the anterior side, here also I don't know how to deal with this side.

I will need to check the image more presicely on a viewer to see what is happening here, but in the case of stenosis, it is possible to lose the anterior part of the canal due to compressions

And a last one, where it's more about the global canal shape.

You can follow here the CSF signal (white), these slices are more tricky since you are at the intervertebral foramina where the spinal rootlets exits,, this is why the R-L boundary is a bit blurry, less clear

[valosekj]

I'm now looking on sub-hamburg02_T2w.nii.gz with @abelsalm, slice 144. GT in light blue, and my suggested annotation in red. I would remove some "anterior" pixels and include more "R-L" pixels. @sandrinebedard what do you think?

[valosekj]

here I see the spinal rootlets (in black) at the edge of the spinal canal, I think I would include them, since we want the entire canal cavity, it is normal that it includes the spinal nerve rootlets.

agree. we want to include all the structures in the spinal canal (i.e., spinal cord in dark, CSF in bright, rootlets in dark)

[sandrinebedard]

I'm now looking on sub-hamburg02_T2w.nii.gz with @abelsalm, slice 144. GT in light blue, and my suggested annotation in red. I would remove some "anterior" pixels and include more "R-L" pixels. @sandrinebedard what do you think?

I agree with you!

[valosekj]

sub-tokyoSkyra03_T2w.nii.gz, slice 141. GT in light blue, and my suggested annotation in red. A similar problem, I would remove some "anterior" pixels and include more "R-L" pixels.

[abelsalm]

Thanks to both of you.

It is more clear now; We spent time with Jan: anyway in most of the cases it is also always a bit ambiguous on some borders of the canal, but it's already way more precise for me on how to correct the segmentations!

I'll update when some will be completed.

[abelsalm]

@valosekj @sandrinebedard What do you think about that ?

https://github.com/user-attachments/assets/3d7188de-1900-436d-8b0a-764e3c0d1385

I wonder if some R/L parts are still a bit under segmented...

[valosekj]

Maybe a few more hyperintense pixels could be segmented on slices 241 (0:01 of the recording) and 178 (0:44)?