• Segments bones from any CT scan
• Based on the state-of-the-art nnUNetv2 package
• 3 models with increasing numbers of labels (17, 38 or 60 labels), including a label for high-radiodensity objects (e.g. surgically implanted hardware)
• Subsegmentation model to segment bones into cortical and trabecular regions
• Trained using a high quality, manually segmented dataset of elderly multiple myeloma patients with low bone density and osteolytic lesions
• High accuracy; average Dice for medium model is 0.935
• GPU, CPU and MPS compute devices supported
• GPU with at least 8GB of RAM is recommended, but will work with CPU only
• Developed and tested on Ubuntu and on Windows using Ubuntu under WSL, but should work on Windows, Linux and macOS

• September 2024: added RAM and GPU RAM estimates to log file
• July 2024: added fix for handling data in non-LPS orientation
• April 2024: updated models and code to nnUNetv2
• October 2023: initial release using nnUNetv1

For more information and if you use Skellytour in your work, please cite our upcoming paper: Mann D.C., Rutherford M., Farmer P., Eichhorn J., Palot Manzil F.F., Wardell C.P. (2024). Skellytour: Automated Skeleton Segmentation from Whole-Body CT Images

• Introduction
• Installation
• Usage
• Available Models
• Description of Output
• Getting Help
• Label List and Description

Segmenting bones from CT scans is required for a range of both clinical and research applications; e.g. diagnosing fractures, surgical planning, producing 3D models of bones, quantitative image analysis. However, segmentation is a time-consuming and laborious process. We present Skellytour, an easy-to-use tool for bone segmentation from CT scans. Skellytour is based on the state-of-the-art nnU-Net and was trained using a high quality, manually segmented dataset of elderly patients with low bone density and osteolytic lesions. Segmentations were verified by a board-certified radiologist and a board-certified nuclear medicine physician. Tested against an internal and two external datasets, the medium model achieved high metrics (Dice scores of 0.935, 0.936, 0.953 and Normalized Surface Distance of 0.993, 0.999, 0.990). It includes 3 models with increasing numbers of labels (17, 38 or 60 labels), with a separate label for high-radiodensity artifacts such as surgically implanted hardware. An optional model subsegments bones into cortical and trabecular regions.

It is recommended but not required that you install Skellytour in a Conda environment using Python v3.9. You can find Conda installation instructions here.

## Create new Conda environment and activate it
conda create -y -n skellytour python=3.9
conda activate skellytour

Download and install Skellytour via GitHub using pip. All dependencies will automatically be installed:

git clone https://github.com/cpwardell/Skellytour.git
cd Skellytour/
python -m pip install .

You can test if Skellytour has installed correctly by running skellytour on the command line. It will print the help message and notify you if a GPU is detected and how many are available.

$ skellytour
error: the following arguments are required: -i
usage: skellytour [-h] -i I [-o O] [-m {low,medium,high}] [-c C] [-d {gpu,cpu,mps}] [-g G] [--overwrite] [--nopp] [--subseg] [--fast]

Skellytour: Bone Segmentation from CT scans

optional arguments:
  -h, --help            show this help message and exit
  -i I                  path to input NIfTI file (default: None)
  -o O                  path to output directory (default: .)
  -m {low,medium,high}  model to use; can be low (17 labels), medium (38 labels, default), high (60 labels) (default: medium)
  -c C                  number of CPU cores to use for preprocessing and postprocessing (default: 6)
  -d {gpu,cpu,mps}      compute device to use (default: gpu)
  -g G                  GPU to use (default: 0)
  --overwrite           overwrite previous results if they exist (default: False)
  --nopp                skip postprocessing on predicted segmentations (default: False)
  --subseg              perform subsegmentation, to predict trabecular and cortical labels (default: False)
  --fast                perform segmentation tasks with a single fold, not the full ensemble model. Not recommended (default: False)

1 GPU detected

Skellytour requires only a gzipped NIfTI file of a CT scan as input. Note that you should provide a path to the actual file, not the directory containing it. The default usage is:

skellytour -i /path/to/input/nifti.nii.gz

This will run the medium model with 38 labels and write output to the current working directory. When a model is run for the first time, Skellytour will download the pretrained model from GitHub and store it in a hidden directory in the current user's home directory e.g. /home/user/.skellytour. Each model requires approximately 1.2 GB of storage space.

Below is a more complex command that would produce a high (60 label) bone segmentation and additional subsegmentation with trabecular and cortical labels, using the second GPU:

skellytour -i /path/to/input/nifti.nii.gz -m high -g 1 --subseg

The following arguments are available:

There are 3 main models and a subsegmentation model. The main models (low,medium, high) have increasing numbers of labels and are detailed in the Label List and Description section of this document. The subsegmentation model runs after the main model if invoked with the --subseg flag and will segment the bones into trabecular and cortical regions.

While running, Skellytour will print messages to the screen updating users on progress. Output files are gzipped NIfTI files containing multilabel segmentations. As an example, the command below will run the medium and subsegmentation models and write output to the outputdir directory:

skellytour -i example.nii.gz -o outputdir --subseg

If you find an issue not covered in this document, or want to request a new feature or model, please open a new issue on GitHub. Note that these models can be quite hungry for RAM and GPU RAM; if you are having trouble, please contact us and we may be able to suggest ways to help or provide a custom model optimized for your use case.

A complete list of all numeric labels in each model is below.