3D Slicer extension for fully-automatic segmentation of CBCT upper airway volumes using "UpperAirwaySegmentator"
UpperAirwaySegmentator is built on the nnUNet framework and segments the pharyngeal and sinonasal airway regions. The dataset was split 70:30 into training and testing sets and included 75 CBCT scans from the University of Alberta, 65 from a private Chilean center, and 80 CT scans (40 pre- and 40 post-operative) from surgical departments in France.
If you use UpperAirwaySegmentator for your work, please also cite nnU-Net:
Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021;18(2):203-211. doi:10.1038/s41592-020-01008-z
This extension is compatible with the 3D Slicer Preview Release (version 5.7.0, or later), downloadable from the official website.
The plugin can be installed in Slicer using the extension manager. It can be found using the search bar by typing "UpperAirwaySegmentator".
After the install process and restart of Slicer, the extension can be found in the module file explorer under Segmentation>UpperAirwaySegmentator.
It can also be found by using the find module button and searching for the keyword UpperAirwaySegmentator.
To use the extension, load a dental CT or CBCT by either drag and dropping the data in 3D Slicer or by using the
DATA or DCM load buttons.
After loading the data, the data will be displayed in the 2D views.
Switch module to the UpperAirwaySegmentator module and select the volume in the first drop down menu.
Choose the trained AI model to use from the AI model drop down menu.
Click on the Apply button to start the segmentation.
During the first launch, the module's dependencies will be installed. These dependencies include :
- The AI model weights
- Light the torch
- PyTorch
- nnUNet V2
After the install, the volume will be transferred and sent to the nnUNet V2 library for processing. If your device doesn't include CUDA, the processing may be very long and a dialog will ask for confirmation before starting the segmentation process.
During execution, the processing can be canceled using the Stop button.
The progress will be reported in the console logs.
After the segmentation process has run, the segmentation will be loaded into the application.
The segmentation results can be modified using the Segment Editor tools.
The segmentation can be exported as STL, NIfTI and/or OBJ using the Export segmentation menu and selecting the export format(s).
The Surface smoothing slider allows to change the 3D view surface smoothing algorithm.
Due to an ongoing issue with Mac devices and Pytorch, GPU acceleration is not available for now on those devices.
On Linux or WSL system, the inference can get stuck at the stage : "done with volume". This indicates that the process has run out of memory. 32 GB of RAM is recommended to run this extension. If you run into this issue, you can create a SWAP file on SSD with 16 GB which should solve the problem.
On Windows, the torch CUDA dependency may not properly be detected and installed.
To solve this problem, you can use the PyTorch Utils extension, uninstall the version of PyTorch and install
a new version of PyTorch by setting the Computation backend compatible with your hardware. Make sure to restart Slicer after performing the re-installation.
The PyTorch version should be greater than 2.0.0 for nnUNet compatibility.
If the weights are not correctly installed, you can install them manually. To do so, go to https://github.com/alejandro-matos/SlicerUpperAirwaySegmentator and select the latest release.
Download the latest .zip file from the release.
Navigate to your UpperAirwaySegmentator folder (this folder can be also found in the module finder window).
Unzip the weight file in the model-specific folder under UpperAirwaySegmentator\Resources\ML.
For example, the default model is stored in UpperAirwaySegmentator\Resources\ML\upper_airway_155cbct.
Create a download_info.json file containing the path to the downloaded zip file for future reference :
{ "download_url": "https://github.com/alejandro-matos/SlicerUpperAirwaySegmentator/releases/download/v1.0.0-alpha/Dataset014_Airways_155CBCT_fold_all.zip " }
Selectable models are defined in UpperAirwaySegmentator/UpperAirwaySegmentatorLib/ModelDefinition.py.
Add one ModelDefinition per trained model:
ModelDefinition(
key="upper_airway_second_model",
display_name="Upper airway second model",
release_asset_name="DatasetXXX_YourSecondModel.zip",
)The key becomes the local weights folder name under Resources/ML.
The display_name is shown in the Slicer module.
Use release_asset_name when the weights are attached to a GitHub release, or download_url when you want to point directly to a .zip file.
Error reads "Failed to load segmentation. Something went wrong during nnUNet processing. Please check the logs for potential errors and contact the library maintainers." Check that the Torch version is at least 2.0.0. Otherwise, try uninstalling the PyTorch extension, restart 3D Slicer, and then reinstall the PyTorch extension again.
This project welcomes contributions. If you want more information about how you can contribute, please refer to the CONTRIBUTING.md file.
Authors: A. Matos Camarillo (University of Alberta), S. Capenakas-Gianoni (University of Alberta), K. Punithakumar (University of Alberta), M. Lagravere-Vich (University of Alberta)
This project references the structural framework of the Slicer DentalSegmentator project extension, licensed under the Apache 2.0 License. Our implementation features independently trained models and custom modifications while maintaining attribution to the original structural framework.
For clinical support: Dr. Silvia Gianoni-Capenakas - capenaka@ualberta.ca, Dr. Manuel Lagravere-Vich - manuel@ualberta.ca
For technical inquiries: Alejandro Matos Camarillo - amatos@ualberta.ca