A large number of AI segmentation models have been developed over the past few years, but TotalSegmentator stands out in several aspects:
it can segment many structures: 104 anatomical structures (all abdominal organs, bones, larger vessels, muscles)
it is very robust: it can segment any whole-body, abdominal, chest CT images, regardless of image resolution and field of view
fast: computation time at full resolution is 1-2 minutes on a CUDA-capable GPU; and about 1-minute at low-resolution mode on CPU
While it is not perfect (there can be a couple of few-millimeter segmentation errors), the segmentation is accurate enough for most purposes - 3D visualization, quantification, specifying region of interest for segmentation, registration, or further analysis.
TotalSegmentator extension can be installed by a few clicks in the extensions manager. It does not require a GPU, it can segment a whole-body CT in about a minute using just the CPU, but a CUDA-capable GPU is recommended for full-resolution segmentation (which takes 1-2 minutes on GPU but it would take 40-50 minutes on CPU).
Demo and tutorial video:
The TotalSegmentator segmentation engine used in this extension was created at the department of Research and Analysis at University Hospital Basel using the nnU-Net framework developed at DKFZ. If you use this tool then please cite: Wasserthal J., Meyer M., Breit H., Cyriac J., Yang S., Segeroth M. TotalSegmentator: robust segmentation of 104 anatomical structures in CT images, 2022. https://arxiv.org/abs/2208.05868. arXiv: 2208.05868
Dear Andras, as I have said in an other post where you replied to a question, you have done an incredible job and I received it as my prefered X’mas gift.
Thanks again and season greeting.
I’ve been following and using Slicer for a while. Today I’m giving a try to this magnific new tool. It really works well with the Sample data, but I get a wrong segmentation when working with my own DICOM file.
I add the file, then open TotalSegmentator with the configs you can see there. What can be happening? Thanks again for you huge contribution for this field.
That result looks reasonable. There’s a big segment called ‘face’ at the front of the head. My understanding is that this can be used to remove features that might otherwise identify the individual subject. Removing these features makes it easier to share 3D data.
I guess that CMF structures can’t be segmented due to data protection on the training process. I’m right?
After reading your comment I have loaded different body CT scans and the result was good.
Right now there are no CMF-specific substructures included in the TotalSegmentator model. There are dedicated models in development to subdivide the head into relevant anatomical regions. Some of that work will be discussed during the upcoming Project Week.
