UltraBones100k is, to our knowledge, the largest dataset of ex-vivo B-mode ultrasound images of human lower limbs with bone-surface annotations — approximately 100,000 annotated ultrasound frames across 14 cadaveric specimens, covering the tibia, fibula, and foot. Labels are generated by an automated pipeline that superimposes a tracked 3D bone model (segmented from preoperative CT) onto tracked ultrasound frames, followed by an intensity-based per-frame refinement that accounts for ultrasound physics. A model trained on this dataset consistently outperforms manual labeling, especially in low-intensity regions (e.g., +320% completeness, +27.4% accuracy, +197% F1 at a 0.5 mm distance threshold).

The dataset supports research and clinical translation of ultrasound in computer-assisted interventions, including 2D bone segmentation, 3D bone surface reconstruction, and multi-modality bone registration (US↔CT).

📄 Paper: Wu et al., Computers in Biology and Medicine 194 (2025) 110435 — DOI
💻 Code: github.com/luohwu/UltraBones100k

Dataset summary


Specimens	14 ex-vivo human lower limbs (`specimen01` … `specimen14`)
Anatomies	tibia, fibula, foot
Modality	Tracked B-mode ultrasound (grayscale PNG) + CT-derived 3D bone models
Annotated frames	>100,000 ultrasound images
Label types	visible bone surface, full bone surface (incl. acoustic shadow), model predictions
Reference geometry	CT bone segmentations (STL meshes), tracking poses, 3D point clouds
Distribution	One ZIP archive per specimen (~1.3–6.1 GB each, ~40 GB total)
License	CC BY 4.0

Files & structure

The dataset is distributed as 14 ZIP archives, one per specimen (specimen01.zip … specimen14.zip). Each archive expands to the following structure:

specimenNN/
├── CT_bone_segmentations/          # Ground-truth bone surfaces segmented from preoperative CT
│   ├── CT_bone_model_merged.stl    #   all target bones merged into one mesh
│   ├── tibia.stl
│   ├── fibula.stl
│   └── foot.stl
├── pretrained_model/
│   └── epoch_30.pth                # Specimen-specific segmentation model (leave-one-out)
└── ultrasound_records/
    └── <anatomy>/                  # fibula | foot | tibia
        └── recordNN/
            ├── UltrasoundImages/   # B-mode frames, grayscale PNG; filename = frame timestamp (e.g. 24363.png)
            ├── Labels/             # Bone-surface masks (directly visible bone), one per frame
            ├── Labels_full/        # Full bone-surface masks, including acoustic-shadow regions
            ├── Labels_pred/        # Model-predicted masks
            ├── 3D_reconstructions/
            │   ├── with_GT_labels/    # point clouds from ground-truth labels (.xyz)
            │   └── with_pred_labels/  # point clouds from predicted labels (.xyz)
            └── tracking.csv        # Per-frame tracking pose for each ultrasound frame

Notes on the contents

UltrasoundImages / Labels / Labels_full / Labels_pred are single-channel PNGs that share the same resolution and filename per frame (the filename is the acquisition timestamp), so a frame and its masks line up by name.
Labels marks the bone surface that is directly visible in the image; Labels_full additionally includes the bone surface that falls within the acoustic shadow (occluded but geometrically present).
3D_reconstructions (.xyz point clouds) are provided for both ground-truth and predicted labels. Filenames follow the pattern reconstruction_pcd[_filtered][_optimizedPose].xyz, where _filtered removes outliers and _optimizedPose uses the intensity-refined frame poses.

`tracking.csv` columns

One row per ultrasound frame, indexed by timestamp (matches the PNG filename).

Column	Description
`timestamp`	Frame id (matches the PNG filename)
`space_factor`	Pixel spacing / scale factor (mm per pixel)
`x`, `y`, `z`	Probe position (initial tracked pose)
`euler_x`, `euler_y`, `euler_z`	Probe orientation (Euler angles, initial pose)
`x_optimized`, `y_optimized`, `z_optimized`	Probe position after intensity-based refinement
`euler_x_optimized`, `euler_y_optimized`, `euler_z_optimized`	Probe orientation after refinement
`width`, `height`	Image dimensions in pixels

Usage

Download a single specimen

from huggingface_hub import hf_hub_download

zip_path = hf_hub_download(
    repo_id="luohwu/UltraBones100k",
    filename="specimen01.zip",
    repo_type="dataset",
)
print(zip_path)

Download the whole dataset

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="luohwu/UltraBones100k",
    repo_type="dataset",
    local_dir="UltraBones100k",
)

Or with the CLI:

hf download luohwu/UltraBones100k --repo-type dataset --local-dir UltraBones100k

Unzip a specimen

import zipfile
with zipfile.ZipFile("specimen01.zip") as z:
    z.extractall("UltraBones100k_extracted")

Citation

If you use UltraBones100k, please cite:

@article{wu2025ultrabones100k,
  title   = {UltraBones100k: A reliable automated labeling method and large-scale
             dataset for ultrasound-based bone surface extraction},
  author  = {Wu, Luohong and Cavalcanti, Nicola A. and Seibold, Matthias and
             Loggia, Giuseppe and Reissner, Lisa and Hein, Jonas and Beeler, Silvan and
             Vieh{\"o}fer, Arnd and Wirth, Stephan and Calvet, Lilian and F{\"u}rnstahl, Philipp},
  journal = {Computers in Biology and Medicine},
  volume  = {194},
  pages   = {110435},
  year    = {2025},
  doi     = {10.1016/j.compbiomed.2025.110435}
}

License

Released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are free to share and adapt the material for any purpose, provided you give appropriate credit.

Downloads last month: 113

Total file size:

41.6 GB

UltraBones100k