How to use nnUNet weights other then TotalSegmentator weights?

[amitjc]

Dear Experts,

Congratulations and sincere thanks for releasing this extension.

I request inputs to run nnUNet models, available at https://zenodo.org/records/3734294, other than the TotalSegmentator nnUNet model. I seek the inputs in order to segment liver, liver tumor and hepatic vessels. These have a different folder structure and model weight extension, compared to the ones recommended for using the extension.

Look forward to hearing from you soon.

Best Regards,
Amit.

[Thibault-Pelletier]

Hi @amitjc,

If you want to run nnUNet V2 models, you can have a look at the Slicer NNUNet extension. This extension allows to load a volume, select a model to use for segmentation and run the segmentation.

It was tested with the DentalSegmentator and TotalSegmentator weights and should work for any nnUNet V2 model.

Best,
Thibault

[amitjc]

@Thibault-Pelletier, thank you for your prompt response.

nnUNet v2 uses a slightly different folder structure, <weight_name> and json file; eg., fold_0 to fold_5, no dataset.json and weights under each fold with model_final_checkpoint.model instead of checkpoint_final.pth

Can these weights be leveraged with Slicer nnUNet extension?

Best Regards,
Amit.

[Thibault-Pelletier]

@amitjc the extension is only compatible with nnUNet V2.

If you take a look at recent TotalSegmentator weights, the folder structure looks like the following :

• Dataset730_TotalSegmentatorMRI_part1_organs_495subj
  • nnUNetTrainer_DASegOrd0_NoMirroring__nnUNetPlans__3d_fullres
    • fold_0
      • checkpoint_final.pth
    • dataset.json
    • plans.json

The extension can handle multiple folds for ensemble inference.

The dataset.json and folds folders + weights are required for the extension to work. This expectation is documented here : https://github.com/KitwareMedical/SlicerNNUnet/tree/main?tab=readme-ov-file#expected-weight-folder-structure