Thanks Brandon, the extension was written by Andras Lasso @lassoan as the principal author and maintained by us as a team.
How does someone detect this? I foresee AI models will be like gold in the sense that when you have a gold bar you never know from where it was sourced after melting it
Does someone share this thought?
Yes, people should be very careful about what data they rely on when training their models in order to respect the licensing terms.
This particular term applies specifically to the restricted models for the subtasks that Jakob mentioned. I understand now that you won’t have access to these models unless you specifically request them from the TotalSegmentator site, so it should be clear when this clause is in effect. But people should be aware so they can make appropriate plans if they want to use those restricted models.
3.2 LICENSEE shall not use the results of the use of the SOFTWARE or any IMPROVEMENTS to train an own similar algorithm. For the avoidance of doubt, “results of the use of the SOFTWARE or any IMPROVEMENTS” means any data, information, or insights generated by the SOFTWARE or any IMPROVEMENTS, including but not limited to predictions and segmentations.
Linked from this site: Streamlit
But I also understand now that this term does not apply to the free models available in v1 and v2 of TotalSegmentator.
NOTE: just to empathize this thought I had is not intended for any AI project/researcher in particular, here is the same post from 2 months ago in my LinkedIn
It may be difficult to trace the origin of a trained AI model using technical means. However, there are legal ways to make it very difficult for someone for taking advantage of inappropriately obtained goods (gold, source code, data, etc). Therefore, licensing restrictions can provide meaningful protection.
A post was split to a new topic: Problem running TotalSegmentator in Slicer-5.4
A post was split to a new topic: Failed to get TotalSegmentator package version on macOS
TotalSegmentator v2 can run using version 5.4.0?
TotalSegmentator v2 requires Slicer-5.5.0 - rev32251 or later, as Python packages bundled with Slicer-5.4.0 are too old, which would make installation of the required Python packages too complicated.
We may update Python packages in the Slicer-5.4.1 patch release in a couple of weeks. If that happens then we’ll switch to TotalSegmentator v2.
Nice! This updated version looks great
it looks amazing !!!
is it available or not yet?
If you install 3d Slicer 5.5.0 (preview) and the TotalSegmentator extenson it is available already.
Thanks so much ! But i have this type of results, not so smooth … somebody had the solution ? thanks
You seem to work in CPU mode, where the volumes are downsampled for the segmentation. Invest in a CUDA-enabled GPU to get better results. You could also smooth your segmentations in the segment editor.
Yess i’m actually work with a macbook pro m2 pro.
Apple hardware has no CUDA-enabled GPU unfortunately.
Apple has some hardware acceleration that PyTorch may be able to utilize, so it is worth trying the full-quality mode (disable “fast” option) and see what segmentation quality you can get and with what computation time.
If you don’t want to invest into buying a computer that has a CUDA-capable GPU then you can rent a GPU-equipped virtual machine from Amazon/Microsoft/Google and install&run Slicer there.
If you are a researcher funded by the US government, it is rather easy to get access to a VM desktop with GPU via ACCESS allocation - quite a few of us in the IDC team have been using those VMs exactly for the purpose of testing/using Slicer functionality that requires GPU: ACCESS allocations - IDC User Guide.