The idea is to include at least C1, therefore the proposed methodology could be:
Select images where C1 is fully covered. Typically: T1w MPRAGE, T2 SPACE, MP2RAGE
Run propseg on these images. Propseg's parameters can be tweaked so that the propagation algorithm does not stop when the cord gets bigger (typically at the medulla oblongata). Relevant param: -max-area <float> and possibly -max-deformation <float>
Automatically select a few vertebral levels that are covered by both propseg and contrast-agnostic methods (eg: C3-C5), and estimate a scaling factor to account for the contrast dependency of propseg. See similar investigation: XXX (@Nilser3 can you please cross-ref your post related to this investigation)
Apply the scaling factor to dilate/erode the propseg segmentation
Merge the segmentation from propseg and contrast-agnostic
This could done by fining the z-index where the contrast-agnostic segmentation stops, subtract a few slices, eg: 5, to that index (because the very top part of the segmentation is often partial, ie: does not cover the full extent of the cord)
Add the propseg part above that identified index
QC the whole thing
Re-train model, by inputting also data from many other contrasts, that do not include the whole C1 (so that the model would not learn to not segment C1).
The idea is to include at least C1, therefore the proposed methodology could be:
-max-area <float>and possibly-max-deformation <float>To be done within that repos: https://github.com/sct-pipeline/contrast-agnostic-softseg-spinalcord