search

wasserth / TotalSegmentator

Tool for robust segmentation of >100 important anatomical structures in CT and MR images
Apache License 2.0
1.28k stars 214 forks source link

Totalsegmentator_mri fails for non-whole-body-MRI and contrast enhancement #316

Open Nanex101195 opened 3 weeks ago

Nanex101195 commented 3 weeks ago

I have abdominal T1-VIBE sequences and totalsegmentator always predicts some lung, where never has been any lung visible. Also segmentation of other organs is not really competitive. I tried it with native T1, as well as contrast enhanced T1s. With contrast enhancement of the liver for example, the model fails completely to predict anything.

How can you claim a "sequence independentness" in your arxiv preprint, where segmentation on MRI is not even scanner invariant in most cases?

For CT this might hold up with the absolute HU, but for MRI? In my tests, it didn't work as explained in your paper, to be honest. I think the inhomogeneities of each MRI Scanner, let alone the sequence specific intensity distributions are not really possible to overcome. Can you explain to me, how you were satisfied with the "sequence independent" claim? I can't understand this in any way and want to understand your thought process about this.

wasserth commented 3 weeks ago

Thanks for the feedback. For all the MR images we tried it worked well. Maybe there is some other problem. What is the command you used? Would it be possible to share the dataset so I can try to reproduce the issue?