Failure cases of the `contrast-agnostic` model

[plbenveniste]

Hi !
I don't know if this is of interest to you, but here is a failure case of the contrast-agnostic model ran with sct version git-master-90fb8654fad06eb94fc4d83075b5a1192c22fea4

Tagging @naga-karthik and @sandrinebedard

[plbenveniste]

Another one:

[plbenveniste]

More example can be found in the full QC: #7

[naga-karthik]

Strange! seems like the model is segmenting something outside the spinal cord as well. can you move the whole qc folder (also the data_processed folder) to somewhere on duke so that i can access it and take a look at the segmentations?

[plbenveniste]

As indicated here #7 (comment), the SC segmentations issues came from motion in the images.

QC and segmentation can be accessed here: ~/duke/temp/plben/ms-critical-lesion/dataset_qc

[jcohenadad]

the SC segmentations issues came from motion in the images

I don't see obvious motion artifacts in #6 (comment) or #6 (comment)

[plbenveniste]

the SC segmentations issues came from motion in the images

I don't see obvious motion artifacts in #6 (comment) or #6 (comment)

Isn't this motion?

[jcohenadad]

Isn't this motion?

No, this is just a representation of the spinal cord in the QC report, wherein each axial sub-figure is centered in the spinal cord based on the center of mass of the segmentation. If the segmentation is failing, the center of mass extraction is also failing, yielding this apparent motion of the cord.

Participant motion can be assessed by looking at the sharpness of the image, and in this case it looks pretty sharp

[plbenveniste]

Okay. I did not make myself clear then. I didn't mean motion of the subject being imaged, but motion between two adjacent slices. It can be seen clearly on this GIF.

This causes problem with SC segmentation as it is supposed to be one connected component. I think that the "motion between slices" causes it to fail.

[jcohenadad]

Outch, that's not good. Well, in this case it is either:

• The acquisition was done with two slabs, with patient table moving in between, and stitching of the two slabs failing because (i) patient motion between the two slabs and/or (ii) gradient non-linearity distortion not properly corrected
• Sequential 2D acquisition, with strong patient motion half-way in the acquisition (unlikely)
• Wrong stacking of the slabs. Do you still have the DICOM data?

How often do you see this? If less than 5% I suggest you exclude this participant.

[plbenveniste]

Wrong stacking of the slabs. Do you still have the DICOM data?

I only received the Nifti data. (and no json file associated)

How often do you see this? If less than 5% I suggest you exclude this participant.

Information is stored #7 (comment). I see it in 11 cases out of 241 images (4.5%).