Is there an full body (or lower limb) muscle segmentation model from MRI images? As I understand totalsegmentator is specifically for CT.
1 Like
There are a couple of pre-trained MRI segmentation models for MONAI Auto3DSeg (brain, prostate, etc) and you can probably find some spine segmentation models, but I don’t remember seeing any for lower lung muscles. However, it is almost sure your can train Auto3DSeg or nnUNet model for this quite easily, by manually segment cases (perhaps a few dozen is enough to get started)
1 Like
We have a project underway to try training an MR segmenter based on TotalSegmentator using the SynthSeg approach (not muscles yet though). The preprint paper actually just went up on arxiv:
Would be great to generalize to other segmentation tasks.
5 Likes
This looks interesting @pieper. Is the trained model available publicly?
Not currently, but I’ll check.
1 Like
Andras, @DeepaKrishnaswamy and @Cosmin_Ciausu are working on this. We absolutely will make everything publicly available, but there are some experiments and refinements that we want to do first. It’s great to see the interest. This most definitely adds a lot to the motivation to continue this work!
2 Likes
Hi @lassoan,
Thank you for your interest! As @fedorov said, we are in the process of doing a few more experiments. We are definitely planning on making the model, and the code, publicly available, and will get back to you when we are at that stage.
Deepa and Cosmin
2 Likes
Following up here, TotalSegmentator v2.2 now has support for MR tasks. SlicerTotalSegmentator will have to be updated to specify to use this new version.
3 Likes
There is also MRSegmentator, also based on nnUnet., and trained on 1400+ MRI images from UKBB.
1 Like
If anyone experimented with this model, I would be very interested to hear about your experience!
1 Like
We have been discussing this with @wasserth. I expect that MRI segmentation will be available in Slicer TotalSegmentator extension from tomorrow. I’ll post updates here.
2 Likes
TotalSegmentator extension is updated now to use the latest models, including the new total_mr. The update is available from tomorrow for both the latest Slicer Stable Release (Slicer-5.6) and latest Slicer Preview Release (Slicer-5.7 downloaded 2024-06-01 or later).
Example result for an abdominal MRI (source image: Imaging Data Commons / CMC-CRC-MSB-01262 / MRI Abdomen / 31 AX T1 3D FS 3MIN POST_W):
Slice sweep:
3D Rotation:
5 Likes
Hi @lassoan
This is looking very nice !
I tried on my MRI data set but even without the fast setting checked, the segmentation is not as good as yours.
Is there a preferred protocol or parameter for the segmentations to work better? I am working with T2 HASTE, with 4mm slices on the transverse plane, but this could change. Do you have any recommendations?
Here is an example of segmentation I did just now with the new total_mr (we can see, for example, that the stomach is not well segmented):
Thank you in advance for your answer!
François
4mm slice thickness is very large. Each voxel is a long stick, it is very difficult to create accurate segmentation at such crude resolution.
You may try resampling the image to have isotropic spacing using Crop volume module to see if it improves the results. Let us know is it helped.
Thank you for your answer @lassoan.
I’ll try it and get back to you.
What slice thickness did you use to get the result you showed in the above video ? Maybe you could share details aboute sequences that would work with this new version of total segmentator ?
Reducing slice thickness will induce other anatomical errors. I think, as MRI sequences are already taking 1min30s, with longer MRI movements will change the gastrointestinal anatomy during the acquisition …
Thanks Andras for updating the TotalSegmentator Slicer Extension to the newest TotalSegmentator version!
The problem with many MR images is the much bigger slice thickness compared to CT. This lower data quality results in lower accuracy compared to TotalSegmentator CT. In our training dataset we have many low resolution MR datasets. For these images it is also difficult to create good ground truth annotations. The alternative would be to train only on nice high quality MR images. Then the segmentation accuracy would be better on high res MR images, but the model would not work as robustly on any MR sequence. We decided to prioritise robustness. For many applications this level of accuracy is still fine. For some applications you need higher accuracy. Then some refinement of the initial segmentation is needed.
2 Likes
On the official download page, I still only find the latest stable release (revision 32448) built on 2024-04-05 without the new total_mr segmentation. Am I overseeing something or was there an issue updating the latest release?
Thanks in advance!
@mwirtz this functionality is available in a Slicer extension. The version you see on the download page is the version of the Slicer application. Individual extensions may have more recent builds.
Thanks @fedorov, yes I tried to install the extension in the latest Stable release and total_mr was not appearing. I assumed it might still have something to do with the built version being older than the extension update.
You are not doing anything wrong. I assume you are trying windows.
There is an ongoing issue about extensions on windows build. Total segmentator is one of those affected (no windows build per Cdash. CDash)
Devs are aware of the issue and looking into it, but I do not know when it is going to resolved.
Linux and and Mac seems unaffected from this. If you have access to those you can on those OS.
1 Like