UltraBonesHip — CT–ultrasound dataset (femur & pelvis)

UltraBonesHip is a multi-modal CT–ultrasound dataset of cadaveric human femur and pelvis, shipped with self-contained code that reconstructs 3D ultrasound point clouds from the raw tracked B-mode frames in one command. It accompanies the Medical Image Analysis 2026 paper (see Citation).

Download

The dataset (data and reconstruction code) is packaged as a single archive, UltraBonesHip.zip (~34 GB). Download and extract it — you'll get reconstruct.py, code/, Dockerfile, requirements.txt, and specimen00–specimen04/. Run the reconstruction steps below from the extracted folder.

Dataset summary

• 5 cadaveric specimens (specimen00, specimen01, specimen02, specimen03, specimen04).
• Anatomies: left femur, right femur, pelvis.
• Modalities: tracked B-mode ultrasound sweeps (every frame carries its optically-tracked probe pose) and CT bone-surface segmentations.
• Each anatomy is covered by several ultrasound sweeps (hundreds to ~1,500 frames each); every frame is pose-tagged, so the sweeps reconstruct directly into 3D point clouds in the CT coordinate frame.
• Probe calibration and all reconstruction parameters are baked into reconstruct.py.

To be uploaded

The following will be released here soon:

• Bone segmentation masks — per-frame 2D bone-surface masks for the ultrasound images.
• Optimized probe poses — refined per-frame tracking poses that further improve the ultrasound–CT alignment.

Directory structure

UltraBonesHip/
├── README.md                      # this file
├── reconstruct.py                 # run US -> 3D reconstruction (model auto-downloads from HF)
├── Dockerfile                     # reconstruction environment
├── requirements.txt
├── .dockerignore
├── code/                          # vendored reconstruction engine
│   ├── reconstruction/            #   segmentation, projection, calibration, denoise, merge
│   └── utilities/
├── models/                        # epoch_30_leave_12_out.pth auto-downloaded here on first run
├── reconstructed/                 # OUTPUTS created by reconstruct.py
│   ├── intraoperative/specimenNN_<anatomy>.xyz    # reconstructed US clouds
│   └── preoperative/specimenNN_<anatomy>.stl      # matching CT bone meshes
├── specimen00/
│   ├── CT_bone_segmentations/     # *.stl (left_femur, right_femur, pelvis, ...)
│   └── ultrasound_records/
│       ├── left_femur_axial/
│       ├── left_femur_coronal/
│       ├── right_femur_axial/
│       ├── right_femur_coronal/
│       ├── left_pelvis_axial/
│       ├── left_pelvis_coronal/
│       ├── right_pelvis_axial/
│       └── right_pelvis_coronal/
│           └── recordNN/
│               ├── poses.csv     # per-frame tracked probe poses (x,y,z + euler)
│               └── UltrasoundImages/   # raw B-mode frames (*.png)
└── ...                            # specimen00, specimen01, specimen02, specimen03, specimen04

Each recordNN/ is one ultrasound sweep.

Data format

• CT_bone_segmentations/*.stl — CT bone-surface meshes (left_femur, right_femur, pelvis, and the per-side meshes).
• ultrasound_records/<sweep>/recordNN/
  • poses.csv — one row per frame: image file (file_path), tracked probe pose (x, y, z in mm + euler_x, euler_y, euler_z in degrees), timestamp, image size.
  • UltrasoundImages/*.png — raw B-mode ultrasound frames.

Reconstruct the US point clouds

Docker (recommended)

docker build -t ultrabones-recon .
docker run --gpus all -v "$(pwd)":/data ultrabones-recon      # remove --gpus all to use CPU

Local (Python)

# 1) install a torch/torchvision build matching your CUDA (or CPU) from pytorch.org
# 2) then:
pip install -r requirements.txt
python reconstruct.py                  # all specimens; --specimen specimenNN to limit

On first run the segmentation model (epoch_30_leave_12_out.pth, luohwu/UltraBones100k_segmentation) is downloaded into models/. Outputs are written to reconstructed/: intraoperative/specimenNN_<anatomy>.xyz (US clouds) and preoperative/specimenNN_<anatomy>.stl (CT meshes). For a quick pipeline check add --frame-stride 10 (coarse + fast); omit it for full-resolution clouds — the pelvis in particular needs the full frame density.

Citation

If you use this dataset, please cite:

@article{wu2026neuralbonereg,
  title   = {NeuralBoneReg: An instance-specific label-free point cloud-based method for multi-modal bone surface registration},
  author  = {Wu, Luohong and Seibold, Matthias and Cavalcanti, Nicola A. and Ao, Yunke and Flepp, Roman and Massalimova, Aidana and Calvet, Lilian and F{\"u}rnstahl, Philipp},
  journal = {Medical Image Analysis},
  year    = {2026},
  doi     = {10.1016/j.media.2026.104133}
}