Filter out erroneous Sacrum labelling for thoracic MRI images

[StijnvWijn]

Hi all,

First of all, thanks a lot for this project! I have been using it on my own (MRI) dataset and it seems very robust against different pathologies and anatomies. However, one thing that I did encounter is the fact that sometimes a large segmentation labelled 'Sacrum' replaces many vertebrae in a t-spine image. The 'raw' outputs look good, so I think the issue is mostly in the iterative labeling step, where a small sacrum segment is predicted somewhere in the thoracic IVDs, which causes all the vertebrae below that point to be classified as 'Sacrum', see the images below:
Step 1:

Step 2:

Step 2 Raw:

My question is two-fold:

• Have you seen this issue before? (maybe during Ensure that labeling is reliable when the FOV is only thoracic  #45)
• Would you be open to a PR that fixes this problem? (probably a combination of size/position filtering before the labeling step)

Thanks in advance for your response and time!

[NathanMolinier]

Hey @StijnvWijn, thanks for opening the issue ! I did not have this error before because I mostly work with lumbar or cervical scans but it looks definitely like a bug ! I will try to reproduce it first then I'll implement a fix. I will keep you in touch.

Btw is you sample private ?

[StijnvWijn]

Hi @NathanMolinier,

Thanks for your quick response!
Sadly my dataset is indeed private, so I can't provide you with the data. I have fixed it (with a quick hack) in my fork, so I can give you a working PR as a starting point.

[NathanMolinier]

Hey @StijnvWijn ! Sorry for the delay, I had to deal with a few things before digging more into this issue.

Thanks to your feedback, I’ve reproduced the issue and implemented a fix—very similar to your suggestion in PR #131. I’ve opened a new PR (#132) which resolves the problem on my end. If you have a moment to verify the fix on your side, I’d appreciate it. Also, thank you for your willingness to collaborate; your contribution is highly valued!

[StijnvWijn]

Hi @NathanMolinier
Thanks a lot for your work, I can imagine there are many other things demanding your attention :)

Sadly your implementation does not fully fix the issue. There are a couple of patients for which it now is better, but the others are still affected. The main problem is that the L5-S1 IVD segment is quite large in the step_2_raw output of these remaining patients (1500-3500 voxels of 1x1x1 mm). That is mainly why I also added the superior-inferior check, as I think that those numbers are quite close to what you would expect in a 'real' L5-S1 IVD.

I can't figure out how to edit your PR, but here it is in mine

[NathanMolinier]

@StijnvWijn I indeed missed this second check in your PR ! Could you give another try ? I now check if C2-C3 is the top most disc and L5-S1 is the bottom most disc.

[StijnvWijn]

Hi @NathanMolinier

Thanks a lot, this solves pretty much all the problem patients!
If you want to dive deep, there are still two patients that remain. I had to investigate quite a bit, but I believe the issue is the fact that for these patients have a split segment with value 5, one that is all the way on the top-most IVD and one on the bottom-most IVD. I checked this by using the Center of Gravity of the segments and doing my checks based on that, but I also still had one patient that did not perform well that is fixed with your implementation, so it might just be a tradeoff.

In any case, this new implementation works a lot better already!