Sct_propseg cord poor segmentation results


#1

Dear SCT experts,

I have T2 images that I am trying to segment using version 4.0.0-beta.5. Overall, I get really good segmentation results, however in a number of participants, I encountered some problems.
Firstly, the segmentation did not propagate, and I used -init-centerline viewer option and the problem was solved.
However, due to the low contrast, the segmentation leaks into the CSF and in order to overcome this problem I used the -mask-correction option. I created mask using fsleyes in which I drew at least 8 points around the cord ( I tried to draw a number of mask correction points along the cord- roughly one set per disk and another set for the middle of the vertebrae).
After seeing the results, my impression was that, from one slice to the next, one can have a drastic change in the shape of segmentation. In order to overcome this, I think that I will need to put mask correction points for almost every slice, so that the segmentation does not leak.

What would be your recommendation in order to solve this kind of problem?

I am attaching one example here.

Thank you very much for your kind attention,
Best Regards,
Merve


#2

Hi Merve,
Can you try with sct_deepseg_sc?
Best,
Julien


#3

Hi again,

Thank you very much for the quick reply and helpful suggestion! I tried sct_deepseg_sc (with default options & no brain).
For one of theT2s it worked perfectly. However, for the other one (where the contrast is quite poor) I still have some problems (but much better segmentation compared to the before!).

Best,
Merve


#4

Hi Merve,
Sorry about the delayed response. Glad that results are better with sct_deepseg_sc. At some point there is a minimum contrast that is necessary to robustly separate the cord from the CSF. If you are not using the bottom part of your images though, you don’t need to bother having that part segmented.
Cheers,
Julien