Non-commercial license of weights?

[lassoan]

I've been testing nnInteractive in 3D Slicer and it works very well. Thanks a lot for developing this and releasing it publicly!
This tool has the potential to be highly impactful (perhaps comparable to TotalSegmentator), but the non-commercial (NC) license for the model weights may ruin everything.

Does NC clause mean that if we use this tool to create some segmentations then we are limited in what we can do with that segmentation?

What is the reason for adding the NC clause? If it is because some NC data set was used for its training, then would it be possible to retrain the model without that data set? A potential slight performance degradation would very well worth it if it allowed this tool to be used universally. How difficult would it be for someone to retrain the model with freely usable data?

(This question is moved from LinkedIn to here, because it was not discussed there; and I would need to know the answers before I can start advertising this tool to the community.)

[FabianIsensee]

Hi Andras, you’ve raised important questions regarding the licensing and implications of using models trained on non-commercial data. These are not trivial points, and we want to make sure we give them proper consideration.
We’ll discuss this internally and follow up with you once we have more clarity on how best to proceed.

[lassoan]

No pressure, just checking in if there is an update on this. We are still very interested in the what you can come up with regarding licensing.

[mariamonzon]

related about trainng the model from scratch, are the weights and architecture compatible with nnUnetv2 training model?

[lassoan]

I would like to announce the Slicer extension tomorrow. It would be nice to know a bit more about possible licensing limitations and planned next steps to adjust the tone of the announcement.

nnInteractive is very impactful in demonstrating that an interactive segmentation model can be actually useful in practice. However, it is not clear yet whether nnInteractive will be itself the model that gets widely adopted, or the community needs to reproduce the results from open data to create a model that can be used without licensing uncertainties.

[FabianIsensee]

Hey, sorry for taking this long. I just wanted you to know we haven't forgotten about this and are still in the process

@mariamonzon nnU-Net was used to train nnInteractive but required a lot of modifications to support the online interactive prompt simulation. So no the model cannot be fine tuned with just nnU-Net

[wasserth]

I could imagine that an "easy" solution would be to retrain the model on only the dataset which have a completely open license. As Andras mentioned, a slight performance degradation would be totally fine in most cases. The most work intensive part is to go through the list of datasets and figure out which ones have a open license and which ones do not. If two or three people do this, it should not be too much work. I would volunteer to do 1/3 or 1/2 of the list. Then retraining the models on only these datasets should not be too much work for @FabianIsensee I guess.
nnInteractive works a lot better than SAM/MedSAM and if you do not limit yourself by the license, then you can become the dominant tool for this application.

[wasserth]

I ran a ChatGPT Deep Research Analysis to find out which dataset allow commercial usage and which ones do not. The results is here:
https://chatgpt.com/s/dr_6835d1ccd0d08191976517ae1074f1af

I manually double checked a few and chatGPT was correct. So i guess the majority is correct. Only roughly 25% of the datasets allow commercial usage. So the big question is how much worse a model trained on only 25% of the datasets will be.

[ValentinKoch]

Is there any progress on this topic?

[sedghi]

interested too

[Nik-VO]

I manually checked the ones that were commercially approved by ChatGPT from the list above. From these, the following ones were wrong and can't be used commercially:

• NSCLC Radiomics (NSCLC Lung CT, TCIA) has a CC BY-NC 3.0 license, so can't be used commercially.
• OASIS states in their license: "OASIS data are provided for non-commercial, academic research purposes only."
• KiTS23 (Kidney Tumor Segmentation 2023) also has a CC BY-NC-SA license, so no commercial use.

[cabinader]

Hello,

I wanted to know if there are any updates regarding the NC clause of the license. Is that something that could change ?

If not, do you plan to release a commercial version of nnInteractive that could allow to derive data which would not be limited by the current license ?

Many thanks for this great work !