Dear SCT team,
I am currently performing segmentation of T1-weighted images of the lumbar spine using SCT.
However, I am having trouble with prop_seg and deep_seg not recognizing the spinal cord area well. In the first place, does SCT support segmentation of lumbar spine MRI anatomy images as well as cervical spine?" I would appreciate your answer.
For lumbar segmentation of T1w images:
- You are correct that the default traditional SC segmentation tools (
sct_deepseg_sc) tend to perform poorly on the lower regions of the spinal cord.
sct_propseg does have quite a few options that allow you to customize the approach: Fixing a failed sct_propseg segmentation - Spinal Cord Toolbox documentation. There may be a set of parameters that would work well for your specific data. But, the parameters will depend on exactly how the segmentation has failed (e.g. over- vs. under-segmentation), so I would need to be able to test on your data to make recommendations.
- We do also have a dedicated lumbar segmentation model (
sct_deepseg -task seg_lumbar_sc_t2w). However, unfortunately it is only available for the T2w contrast at this time.
- SCT is also in the process of working on a contrast-agnostic segmentation model, which may work better on T1w lumbar regions than the traditional sct_propseg/sct_deepseg_sc. I would be happy to update this post once the model is available for testing.
Thank you for your kind reply.
I tried the troubleshooting methods listed in the link you posted (Parameter “-init”, Parameter “-init-mask”, Parameter “-init-centerline”, Smoothing the image), but none of them worked.
As you can see in the attached image, only a small portion of the spinal cord is segmented.
At this point, we think that the quality of the image (T2MEDIC) may be too rough, and we are working on a new imaging protocol that will allow us to get a cleaner image.
Do you have any other ideas on this?
We have recently developed a new model for segmenting the spinal cord, that is more robust across various MRI contrasts. If you send me your image in private, we could give it a try, and if it works well we will give you access to that model before the next public release.
Thank you for your kind suggestion.
When you say “send image in private,” is it OK to send it to your email address, or is there such a function in this forum?