Adding Skeleton-Recall to Existing nnUNetv2 installation

[rohand24]

Hi,

Is it possible to add Skeleton-recall trainer to existing nnUNetv2 installation?

I am able to figure out the skelton-recall trainer in the variants directory of trainers.
If i copy that to my nnUNetv2 installation, will it work out of the box?

Regards,
Rohan Dhamdhere

[mrokuss]

Hi @rohand24,

Thank you for reaching out.

The Skeleton-Recall Trainer is not yet included in the official nnUNetv2 release, but we are currently discussing ways to integrate it in the future.

For now, if you'd like to incorporate it into your existing nnUNetv2 installation, you would need to do more than just copy the nnUNetTrainerSkeletonRecall class. You'd also have to integrate the skeletonization process during data loading, which you can find here, as well as the custom loss function here and the compound loss combination here.

Let me know if you need further assistance!

Best,
Max

[wasserth]

Is there any news on adding the skeleton-recall loss to the official nnUNetv2 repository? This would make this loss a lot more useable. I am thinking about using it in TotalSegmentator but having to support two different nnunet installations is stopping me from using it so far.

[ykirchhoff]

Hi Jakob,

there is a new minor version of nnUNet coming soonish and we plan to put Skeleton Recall in there. We expected that to come earlier, so it delayed a few things unfortunately.
Looking forward to seeing Skeleton Recall in TotalSegmentator!

Best,
Yannick

[ykirchhoff]

But actually now that I think about it, it might be easier to just put it in the current version already, so we don't have to wait longer.

[wasserth]

Great! Looking forward to it.