Get_centerline finds the centerline, but deepseg fails to find the sc in the same slice? 2d kernel better than 3d?

[Daniel_Papp]

Dear SCT experts!

I have found some behaviour I can’t explain to myself, and I need expert advice to figure out what is going on.

Context: we have some relatively poor GRE data (7T) which we want to use for SC segmentation. I have found that sct_deepseg_sc (or propseg) does not find the spinal cord in every slice, but sct_get_centerline is able to find the centerline everyhwere. I don’t understand how one is possible, but not the other.

In depth
I tried segmenting the scans, using T2 contrast, but this segmentation results in some holes, and misses a lot of the thorasic region (not surprising due to the low signal):

sct_deepseg_sc -i GRE.nii -c t2 -qc ./qc -kernel 2d -o seg_2dkernel.nii.gz   

I wanted to fill the holes using the -kernel 3d flag, but that made things worse, not better:

sct_deepseg_sc -i GRE.nii -c t2 -qc ./qc -kernel 3d -o seg_3dkernel.nii.gz         

Then I used sct_get_centerline to, well, get the centerline, that worked like a charm*:

Next, I tried using this centerline as an input to help with the segmentation, but it did not help:

sct_deepseg_sc -i GRE.nii -c t2 -file_centerline GRE_centerline_t2.nii.gz -o seg_with_centerline_t2.nii.gz -qc ./qc    

Red: segmentation using 2d kernel, without centerline input, yellow: segmentation using 2d kernel with centerline input, white: centerline

The same is observed using 3d kernel:

For completeness sake, I repeated this using T2* contrast. All the data is in the attached zip.

I have a workaround for my problem, so I am asking more for the sake of my understanding. I can see that sct_deepseg_sc is “Finding the spinal cord centerline…”, so I don’t understand what is going on.

Is this just a case of PEBCAC? Is our data this bad?

Best

Daniel
Example_4_SCT.zip (3.0 MB)

Software versions:
OS: Mac OS Monterey 12.3.1
SCT: git-master-3cfab1dbff6a1467ca2fad03d070eb2762dfa5e1

[jcohenadad]

Hi Daniel,

Thank you for the well-documented issue report!

Yes, if contrast is too low, the deep learning based method won’t output any mask, which is exactly what you are observing here. My suggestion would be to switch to sct_propseg, which in general is more forgiving to poorly contrasted data.

[Daniel_Papp]

Hi Julien

Thanks, I tried that, it does look better. I don’t fully understand why supplying it a centerline did not help. Is it because centerline is only used to find the area around which the SC might be (“look here, sct_deepseg_sc!”), but the actual segmentation still relies on there being enough signal and contrast to segment it (“well I did look, but I can’t tell anything based on this poor contrast…”)?

And sct_get_centerline can find the centerline in slices where deepseg fails, because it has lower requirements (“not too many voxels away from where it was in the previous slice”)?

[jcohenadad]

I tried a few parameters, but I was not super successful . Here is a QC of a few params search:
qc-julien.zip (1.6 MB)

Note: it looks like switching to “t1” contrast was in fact quite successful, so maybe you could try with that and play with the other params

[jcohenadad]

Is it because centerline is only used to find the area around which the SC might be (“look here, sct_deepseg_sc! ”), but the actual segmentation still relies on there being enough signal and contrast to segment it

yup! that’s it