BoneAGUNet

Atlas-guided erosion segmentation for an already-cropped MCP2 or MCP3 joint image stack. The package interface deliberately excludes cohort discovery, SLURM, and ARC filesystem paths.

Pipeline overview

The input image is converted into an edge mask and a closed bone mask while a healthy atlas is deformably registered to the same anatomy. Their difference identifies erosion candidates. Each candidate is passed to AG-UNet as a three-channel ROI containing the image, bone edges, and registered atlas.

Installation and use

git clone https://github.com/ManskeLab/BoneAGUNet.git
cd BoneAGUNet
pip install .
boneagunet-install ~/.cache/boneagunet
boneagunet -i joint_mcp2.nii.gz -o erosions.nii.gz --mcp 2

Python:

from boneagunet import run
run("joint_mcp3.nii.gz", "erosions.nii.gz", mcp=3)

ANTs (antsRegistration and antsApplyTransforms) must be installed separately. The package installs the pinned Manske Lab nnU-Net fork required by the attention checkpoints.

The input is one unprocessed, already stack-registered 3-D MCP joint image. The pipeline performs soft-tissue stripping, MC/PP masking, edge and closed-edge prediction, atlas registration, candidate extraction, erosion prediction, and recombination into the input image geometry. --keep-work preserves all intermediates for inspection.

Use --modality sr-cbct for SR-CBCT inputs; HR-pQCT is the default.

Atlases and intermediate masks

MCP2 atlas	MCP3 atlas

The healthy representative HR-pQCT atlases are shown as sagittal maximum intensity projections.

Input image	Predicted edge mask	Closed bone mask

The second segmentation model receives both the image and predicted edge mask to complete the cortical surface used during atlas subtraction.

Model assets

The trained models and MCP2/MCP3 atlases are hosted at YousifKhoury/BoneAGUNet and are downloaded automatically or explicitly with boneagunet-install.