USCorUNet is the ultrasound deformation estimation model released with
"Robotic Ultrasound Makes CBCT Alive". It estimates dense bidirectional
deformation fields between adjacent ultrasound frames and is used for
deformation-aware CBCT slice updating in robotic ultrasound-assisted workflows.
This Hugging Face repository stores model checkpoints only. The inference
code and examples are released separately in the project code repository:
github.com/ziyuan-li/robotic-us-cbct-updating.
Model Description
USCorUNet combines:
a ResUNet-style context encoder-decoder operating on adjacent ultrasound
frames and simple gradient cues;
a shared-weight correlation encoder;
local correlation volumes for efficient dense deformation estimation.
For an adjacent ultrasound pair (I0, I1), the model predicts:
F01: forward flow from I0 to I1, defined on the I0 grid;
F10: backward flow from I1 to I0, defined on the I1 grid.
These flows can be used to warp ultrasound frames directly or to propagate a
corresponding CBCT slice across time.
Checkpoints
File
Model
Recommended use
uscorunet_base.pth
USCorUNet-Base
General ultrasound deformation estimation
uscorunet_probe_adapted.pth
USCorUNet-Probe-Adapted
Probe-induced deformation and CBCT updating demos
uscorunet_external_adapted.pth
USCorUNet-External-Adapted
Externally induced deformation
Usage
Download the checkpoints from the code repository with:
python download_checkpoints.py
or download a single file through the Hugging Face Hub:
The checkpoints are PyTorch state dictionaries intended to be loaded with the
USCorUNet architecture provided in the accompanying code release.
Intended Use
These checkpoints are intended for non-commercial academic research on:
ultrasound-based dense deformation estimation;
robotic ultrasound;
deformation-aware CBCT slice updating;
ultrasound-guided image registration;
intraoperative motion visualization.
Limitations
The checkpoints are released for research use only.
They are not approved for clinical, diagnostic, therapeutic, or other medical
decision-making use.
Performance may vary for imaging devices, acquisition protocols, anatomies,
and motion regimes outside those studied in the paper.
Citation
If you use USCorUNet or these checkpoints, please cite:
bibtex
1@article{li2026robotic,
2 title={Robotic Ultrasound Makes CBCT Alive},
3 author={Li, Feng and Li, Ziyuan and Jiang, Zhongliang and Navab, Nassir and Bi, Yuan},
4 journal={arXiv preprint arXiv:2603.10220},
5 year={2026}
6}
License
The model checkpoints are released under the Creative Commons
Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.
Commercial use requires permission from the authors.
