Skip to content

3D Slicer nnUNet integration to streamline usage for nnUNet based AI extensions.

License

Notifications You must be signed in to change notification settings

KitwareMedical/SlicerNNUnet

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

20 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Slicer nnUNet

Table of contents

Introduction

This module allows to install and run nnUNet trained models in 3D Slicer with only the training folder.

It streamlines development and integration of new nnUNet models into the 3D Slicer environment.

Acknowledgments

This module was originally co-financed by the Fédération Française d'Orthodontie (FFO) as part of the Dental Segmentator developments and the Cure Overgrowth Syndromes (COSY) RHU Project.

The installation steps are based on the work done in the Slicer Total Segmentator extension.

This extension interfaces 3D Slicer with the nnUNet library.

Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., & Maier-Hein, K. H. (2021). nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2), 203-211.

Using the extension

This extension can be installed directly using Slicer's extension manager.

Once installed, navigate to Segmentation>nnUNet in the modules drop down menu or search directly for nnUnet.

Once in the widget, you can install nnUNet's dependencies by clicking on the "nnUNet Install" button.

This area will display the current version of nnUNet in 3D Slicer's environment. If nnUNet is not yet installed, the current version will display None. It can be installed by clicking on the install Button.

After the install is complete, the current version should display the latest nnUNet version or the version as set in the "To install" field.

Note that the extension may require a 3D Slicer restart before running the inference.

Once the nnUNet is correctly installed, click on the 'nnUNet Run Settings' button to set the path to the Model to use. This path will be saved for further usage after the first segmentation.

The provided model path should contain the nnUNet 'dataset.json' file.

Note

To test this extension, you can download the Total Segmentator NNUNet weights and use the CTChest Sample.

Warning

The model weight folder structure should follow the nnUNet expected weight structure for this module to work (see the expected weight folder structure section)

Select the volume on which to run the model using the volume input editor.

Then click on the Apply button.

The logs console will display all the information regarding running the input model.

Once the model has finished running, the segmentation will be loaded into 3D Slicer with its associated labels. It can then be viewed and edited in the Segment Editor module.

Expected weight folder structure

To properly run the nnUNet prediction, the nnUNet weight folder structure should be preserved.

nnUNet structure should look like the following :

  • <dataset_id> : The dataset id should be either a string beginning by Dataset or an integer
    • <trainer_name><plan_name><config_name> : default configuration folder name is nnUNetTrainer__nnUNetPlans__3d_fullres
      • fold_<> : Fold folder from 0 to n containing the training weights
        • <weight_name> : By default the weight name is checkpoint_final.pth it can also be modified using the Parameter class
      • dataset.json : Configuration file containing dataset labels

If this structure is not preserved, the inference will raise an error.

The Parameter class provides an isValid method to check if the provided model structure is valid.

Please read the official nnUNet documentation for more information.

Contributing

This project welcomes contributions. It follows Slicer's rules for contributions. If you want more information about how you can contribute, please refer to the CONTRIBUTING.md file.