Manual Segmentation as Ground-Truth


I am looking to evaluate the performance of several segmentation algorithms for spinal MRI (T1, T2, T2*).

Several publications have used manual segmentation as ground-truth to compare to automated segmentations (10.1016/j.neuroimage.2017.03.010, 10.1016/j.neuroimage.2018.09.081).

Whilst I am familiar with SCT (primarily through the excellent tutorials/courses by the SCT group- thank you!), this the first time that I will be analysing my own dataset. I was therefore wondering if I could kindly be guided to any resources regarding manual segmentation of the spinal cord, with the view that I am later able to use my manual segmentations as a gold-standard for comparison to automated segmentation.

Another option would be that perhaps I could upload a couple of manual segmentations and be guided from there?

As a separate question (although related), could someone kindly confirm that in the spinal cord grey matter segmentation challenge (10.1016/j.neuroimage.2017.03.010), all of the algorithms were fully-automated (i.e. manual correction was not allowed)? Apologies if this is a naive question.

Thanks so much for all of the help in advance.


Thank you for reaching out.

Regarding generating the ground truth, you can generate them manually or run an automatic segmentation and correct them. Here is a video on how to correct for segmentations:

Manually segmenting the spinal cord is time consuming, and should ideally be done by multiple raters to remove possible inter-rater biases.

We now have a new model to segment the spinal cord in SCT v6.2 ([2310.15402] Towards contrast-agnostic soft segmentation of the spinal cord), in sct_deepseg -task seg_sc_contrast_agnostic that should work on all of your contrasts!

Regarding the gray matter segmentation challenge, in Table 4 you can see if the methods were automatic or not.



Dear Sandrine,

Thank you so much for your kind response.

That is exceptionally useful!

All the best,