How to measure spinal disc heights

[karllandheer]

Hello, I would like to measure the disc heights (i.e., in the inferior/superior direction, after straightening) from the segmentations provided by totlaspineseg. Is there a way an easy way to do this, or do I have to do this via custom code? Any help would be greatly appreciated.

[NathanMolinier]

Hi @karllandheer,
Thanks for your interest in TotalSpineSeg! Currently, this repository doesn't include scripts for spinal cord straightening or disc height extraction. However, TotalSpineSeg is now integrated into the Spinal Cord Toolbox (SCT), which provides several tools for straightening. Once the scans are straightened using SCT, extracting disc heights should be relatively straightforward. I would be happy to help you do that and provide more explanations about SCT if you want !

Also, I'm actively working on retraining TotalSpineSeg to improve the quality of the segmentation in both the cervical and thoracic regions, so stay tuned, it will help you extract more accurate dimensions of the discs.

[jcohenadad]

Hi @karllandheer, in addition to what @NathanMolinier mentioned, you need to be very careful because while the straightening is fairly accurate in preserving morphometric properties in the vicinity of the spinal cord, it can lead to deformations for objects "far" from the spinal cord centerline, like the discs. This is especially true for large spinal cord curvature. Just something to keep in mind.

[karllandheer]

Hello, thank you very much for your response and great work, I attended your talk at ISMRM and it was excellent. Do you have a rough timeline on the re-training of TotalSpineSeg? Regarding the straightening, do you think it's best to use the centerline from the spinal cord (or probably spinal canal, because the spinal cord seems to end around T12-L1), or by fitting the centroid of the discs or vertebrae's to a spline and using that? This alleviates the issue of being "far" from the spinal cord centerline, but then introduces the issue of potentially a noisier centerline

[NathanMolinier]

It is difficult for me to precisely tell when I will have a new model because I'm just beginning the manual corrections of the ground truths (maybe 2-3 months or less).

Regarding the straightening these are really good questions ! I guess choosing between the cord or the canal really depend on the field of view but also right now the cord segmentation is still more accurate than the canal. Using the discs for registration could be an excellent idea too but it is difficult to tell in advance without comparison.

[karllandheer]

OK, thank you. I will keep this in mind, on a related note, I have run this on N=1 UKB subjects, and it took about 2 hours for inference on a CPU. My guess is this is because its interpolated to 1mm^3 isotropic, whereas UKB whole-body Dixon is natively about 2.2 x 2.2 x 3 mm or so. The resulting 3D matrix is thus very large. If there was an option to speed this up it would be great. Something like removing test time augmentation or ensembling would be great. The issue is that running a 2-hour algorithm across all N=80K+ subjects can be quite expensive. I understand if this is not a priority as most people do not have such large datasets

[NathanMolinier]

Unfortunately, the resampling to 1mm^3 is needed because the model was only trained for this specific resolution. That's unfortunate that you don't have access to a GPU. Regarding test time augmentation maybe I could add a flag like nnunet does --disable_tta but I don't know how different the images would look and how much time we can gain. Would you like to have this flag ? Speeding the inference is definitely something that I want to do !

[karllandheer]

For me a slight reduction in performance at the cost of a speedup would be beneficial. From my experience tta has very little impact anyways, yes that would be great. Also I do have access to a GPU, I haven't benchmarked it to CPU yet, but although GPUs are much faster they're also much more expensive, so I have to see which is cheaper in the end

[jcohenadad]

@karllandheer as we are also working on an improved version of the model, how about we join forces and work together on an updated version of the model that also includes UKB data?

[karllandheer]

Hello, yes that would be great. I talked to Nathan at ISMRM about this, we are interested in using this on UKB. Please let me know if I can help with anything.

[jcohenadad]

ok nevermind-- I thought that you were planning to retrain the model-- now I understand that it is on @NathanMolinier 's court to do this. Happy to collaborate either way (eg: for validating new model on UKB). We will keep in touch

[NathanMolinier]

@karllandheer I just looked at the test time augmentation (TTA) from nnunet and saw that only mirroring is used for TTA and because we don't use mirroring during training I believe the TTA is already disabled. To double check this, I created a branch nm/add_disable_tta_flag and got the same computation with and without the flag --disable-tta. So unfortunately TTA was not extending the inference time :/. However, I will let you know if I manage to speed up the inference time !

[karllandheer]

OK. Regarding training on UKB, I was not planning on re-training the model on UKB. I may be able to provide some annotations that you guys could use to train, although sharing UKB data is tricky. I will look into this and get back to you.