UltraBoneUDF: Self-Supervised Bone Surface Reconstruction from Ultrasound Using Neural Unsigned Distance Functions

---

This repository contains the official implementation of UltraBoneUDF.

Paper: https://doi.org/10.1016/j.compmedimag.2025.102690

Updates

2026-10-09, point-to-plane Neural-Pull target. The reconstruction used to stop at the border of the scanned region and leave small holes, so part of the bone surface near that border was never covered. A small change to the Neural-Pull target reduces this loss of coverage. On UltraBones100k the GT-to-reconstruction distance drops from 1.85 mm to 1.15 mm on all 28 cases. Set nc_plane_mm = 2.0 in the conf file; 0 keeps the original behaviour. The best value depends on the data, so compare the two on your own point clouds. Section 7 explains the problem and the change: The edge effect and the point-to-plane target.

1. Environment

The code has been tested with the following setup:

• Operating System: Windows 11 and Ubuntu
• Python: 3.10, 3.11
• CUDA: 12.6, 12.8
• PyTorch: 2.7.1

To run the code, you can either:
a) install the dependencies, or
b) run the project using the provided Docker image.

1.1 Install dependencies

Install all packages listed in requirements.txt:

pip install -r requirements.txt

1.2 Docker

A Dockerfile is provided in the root directory. Alternatively, you can use the prebuilt Docker image directly:

• luohwu123/ultraboneudf:latest

2. Dataset

You can download the UltraBones100k dataset here: https://github.com/luohwu/UltraBones100k

The dataset path is defined in conf/conf_UltraBones100k.conf (default: ./data/UltraBones100k/). If your dataset is stored elsewhere, update this path accordingly.

The dataset includes point clouds reconstructed from freehand ultrasound scans (.xyz files). By default, we use the versions constructed using the original tracking data and predicted labels. If you would like to use other versions (or your own dataset), modify models/dataset.py accordingly.

Note: The provided point clouds are not preprocessed. Applying additional preprocessing, such as outlier removal, can improve performance. For example, statistical filtering can reduce the Hausdorff distance.

3. Train and evaluate on UltraBones100k

To reconstruct bone surfaces on UltraBones100k, run:

python main.py --mode train

Depending on your GPU, reconstructing a single point cloud typically takes under 5 minutes. After each reconstruction, the Chamfer distance is computed and printed. Outputs are saved under:

• outs/UltraBones100k

Alternatively, to load and evaluate using the pretrained model:

python main.py --mode validate

4. Double-face rendering for unsigned distance fields

As noted by the authors of DualMesh-UDF:

“When dealing with UDF fields, there is no theoretical guarantee for extracting a manifold mesh.”

Therefore, reconstruction results should be rendered in double-face mode (or post-processed accordingly). More details can be found here: cong-yi/DualMesh-UDF#2 (comment)

In MeshLab, this can be enabled via the option highlighted by the red arrow in the screenshot below:

5. Reference

@article{WU2025102690,
title = {UltraBoneUDF: Self-supervised bone surface reconstruction from ultrasound based on neural unsigned distance functions},
journal = {Computerized Medical Imaging and Graphics},
pages = {102690},
year = {2025},
issn = {0895-6111},
doi = {https://doi.org/10.1016/j.compmedimag.2025.102690},
url = {https://www.sciencedirect.com/science/article/pii/S0895611125001995},
author = {Luohong Wu and Matthias Seibold and Nicola A. Cavalcanti and Giuseppe Loggia and Lisa Reissner and Bastian Sigrist and Jonas Hein and Lilian Calvet and Arnd Viehöfer and Philipp Fürnstahl},
keywords = {Ultrasound bone surface reconstruction, Neural unsigned distance function, Neural implicit representation},
}

6. Code reference.

This code is adapted from the following repositories (many thanks to the contributors):

• FUNSR: https://github.com/chenhbo/FUNSR
• DualMesh-UDF: https://github.com/cong-yi/DualMesh-UDF

7. The edge effect and the point-to-plane target

UltraBoneUDF reconstructs an open surface and adds nothing where there is no data, so the surface ends at the last input point. The bone within a few millimetres of the border of the scanned region, or inside sparse holes, stays uncovered: on UltraBones100k the GT-to-reconstruction distance is 4.9 mm within 2 mm of that border and 0.7 mm in the interior. The cause is the Neural-Pull target, not the mesher. A query just past the border is matched to a border point, so the field learns a positive distance there and its zero set stops at the data.

The update changes that target from point-to-point to point-to-plane, as in ICP. When the sideways offset between a query and its matched input point is below nc_plane_mm (2 mm), the network learns the distance to that point's tangent plane, using the normals the dataset already estimates; for larger offsets the original target applies. The surface can then continue across the border and across holes for up to 2 mm. No loss term and no data are added, and the mesher is unchanged. With nc_plane_mm = 0 the code runs exactly as published. We chose 2 mm after testing 1, 2 and 3 mm: at 1 mm almost nothing changes, at 3 mm the mesh gains too much surface and the reconstruction-to-GT distance rises. The value depends on the data: it trades coverage against extra surface near clutter, so for your own point clouds compare nc_plane_mm = 0 with nc_plane_mm = 2.0 and keep the one that works better.

License

This work is licensed under the Creative Commons Attribution 4.0 International License. Details in license.txt

Questions or Feedback?

If you have questions, you can open a new GitHub issue within this repository, and we'll get back to you!