What happens when encountering transitional vertebrae when segmenting the spine in TotalSegmentator?
From vert519
How to handle this situation? Who has good advice?
From vert519
There is no automated method that works perfectly all the time. Even if the algorithm works perfectly for 99% of the cases, you still need to allow users to perform manually corrections. In medical imaging, 95% is a commonly accepted success rate, so manual corrections/overrides are very important.
If you have a segmentation step in your workflow then you normally allow users to interactively inspect segmentation results and make manual correction as needed. You can simply add a Segment Editor widget in a collapsible section, or you can created a simplified GUI (for example, a few buttons and maybe a segment selector that allow users to hightlight a region in 3D using Scissors tool and move the designated region to another segment).
1 Like
I try to segmente 10 spinal CT scans , but the results were not very satisfactory Nearly half of the ‘transitional vertebrae?’… Is it a ‘transfer vertebra’ or a segmentation error…? 5 of 10 cases got the result … Maybe I did something wrong?
: Pay attention to adjacent vertebrae with incomplete segmentation of the same color . Is it a ‘transfer vertebra’ or a segmentation error…? 5 of 10 cases got the result … Maybe I did something wrong?
@lassoan@pieper@Juicy@jamesobutler @jcfr
(https://drive.google.com/file/d/1hAZHCWXlASQdbdX23-9re57IMfElmmyI/view?usp=sharing)It includes the original CT and segmentation results. Which master can help to try segmentation and compare the segmentation results?
包含了原始的CT和分割结果, 哪位老师帮帮忙尝试分割一下, 并进一步比较分割的结果呢?
Transitional vertebra (L6, T13) are not labeled by the TotalSegmentator model (Spine segmentation format · Issue #88 · wasserth/TotalSegmentator · GitHub). If you encounter such patient cases then you can fix the segmentation using the Scissors effect in Segment Editor module (and do final touch-up using Paint effect, if needed).
5 of 10 cases(which may be a special case) got this result👆🏻 , well… this ttSeg model is not suitable for fully automated spine segmentation? Are there any other models specifically designed for spine segmentation? @lassoan
Prevalence of these anatomic variations are not 50%, but rather about 10%. If you have incorrect some segmentation for 50% of the cases then probably there are some other issues.
TotalSegmentator is very robust and readily usable for getting a general anatomic context. For example, if you do lung surgery, then TotalSegmentator’s spine segmentation is perfectly suitable as is. If you do spine surgery then TotalSegmentator can venerally provide a good initial segmentation (to approximate bounding box for each vertebra) that you can use as input for specialized spine segmentation model. Specialized spine segmentation models are being developed, for example by the team of @Ron.