search

ivadomed / model-spinal-rootlets

Deep-learning based segmentation of the spinal nerve rootlets
5 stars 2 forks source link

Explore skeletonization filters (or similar approaches) to extract the centerline of nerve rootlets segmentation #59

Open jcohenadad opened 1 month ago

maxradx commented 3 weeks ago

This issue is about exploration of skeletonization filters (or similar approaches) to extract the centerline of nerve rootlets segmentation.

An assumption is that nerve rootlets are tubular structures on medical images, so filtering approaches should work on tubular structures.

1- Multiple filtering modules exist in 3D Slicer. Examples from Slicer documentation here.

2- It is possible to extract skeleton of a binary object (link here). However, there does not seem to have default 3D configurations. Would it work on 3D structures?

3- An extension from which the centerline can be extracted from tubular structure is VMTK (Vascular Modeling Toolkit). Initially designed for vascular structures, it is claimed to work on any tubular structures (like vessels, nerve rootlets ...) A potential user should consider that the last commit on the main branch of github has been 2 months ago (from June 24th, 2024). Github repository here. Moreoever, when the keyword 'centerline' is used in the 3D Slicer extension manager, only SlicerVMTK is in the results.

image

4- Retrospectively, I asked chatGPT 4-o the question about exploring skeletonization filters to extract the centerline of nerve rootlets segmentation. Although there might have some errors, I found its answer useful for providing a way of how performing centerline extraction from nerve rootlets in 3D Slicer (joint text file).

20240624_chatgpt_answer.txt

jcohenadad commented 1 week ago

The tubular extension from Slicer3D seems to work quite well, and is being used for labeling lumbar rootlets.

We might want to consider re-thinking ground truth labeling method for cervical rootlets as well, but that would imply re-doing all GTs, which is a big task...

valosekj commented 1 week ago

I agree that the tubular extension is great for lumbar rootlets, especially when creating them from scratch.

For the cervical rootlets, the situation is slightly different: in most cases, we get a relatively good initial prediction and only need to correct a few slices, sometimes only a few pixels. This can be done relatively quickly even with FSLeyes/ITKsnap. But agree, we should keep the tubular extension in mind even for cervical rootlets. Let's see how many corrections we need for MP2RAGE images (https://github.com/ivadomed/model-spinal-rootlets/issues/63#issuecomment-2218257267).