BoneFM

BoneFM

BoneFM is the skeleton-focused CT foundation backbone used by BoneCoT: Multi-center validation of a whole-body skeleton foundation model for bone metastases guided by clinician-derived chain of thought.

• Paper DOI: 10.1038/s41551-026-01736-1
• Code: FrankZhangRp/BoneCoT
• Project page: frankzhangrp.github.io/BoneCoT

Paper and Authors

Hui Zhao^1,,#, Ruipeng Zhang^2,, Zhiyu Wang^1,*, Yifeng Gu², Shengyuan Xu³, Sheng Wang^4,#, Yuehua Li^2,#

1. Metastatic Bone Tumor Clinical Center, Shanghai Sixth People's Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, China
2. Institute of Diagnostic and Interventional Radiology, Shanghai Sixth People's Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, China
3. Mailman School of Public Health, Columbia University, New York, NY, USA
4. Paul G. Allen School of Computer Science and Engineering, University of Washington, Seattle, WA, USA

^*These authors contributed equally: Hui Zhao, Ruipeng Zhang, Zhiyu Wang

^#Email: [email protected]; [email protected]; [email protected]

Model Summary

BoneFM is a Vision Transformer backbone adapted from DINOv2-style self-supervised learning for skeleton-focused CT representation. BoneCoT uses BoneFM features and clinician-derived task dependencies for downstream bone metastasis and bone-related disease reasoning.

This repository hosts the public BoneFM backbone checkpoint:

Checkpoint integrity:

• Size: 4,946,789,774 bytes
• SHA256: 5bed7f117e4f8a9f3b11eded0408e9ba60ee0bf3c3b335982d6b9e608c69d271

Intended Use

BoneFM is intended for non-commercial research on skeletal CT representation learning and downstream bone-related disease modelling. It can be used as a feature backbone with the public BoneCoT code when users provide their own de-identified image data and clinically appropriate labels.

BoneFM and BoneCoT are not standalone clinical diagnostic devices. They should not be used for patient management without local validation, regulatory review, and qualified clinical oversight.

Input Convention

Prepare CT slices with the bone-window convention used by the public BoneCoT code:

WL = 300
WW = 1500
image = clip((HU - (WL - WW / 2)) / WW, 0, 1)

The public code expects PIL-readable RGB-compatible image files and applies the evaluation transforms defined in the BoneCoT repository.

Download

Download the released checkpoint:

hf download frankzhang/BoneFM BoneFM.pth --local-dir finetune/checkpoints

Python alternative:

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="frankzhang/BoneFM",
    filename="BoneFM.pth",
    local_dir="finetune/checkpoints",
)
print(path)

The expected local path for the BoneCoT repository is:

finetune/checkpoints/BoneFM.pth

Public Release Boundary

This model repository is for the BoneFM backbone checkpoint and model card. It does not include:

• Private clinical training, validation, or test datasets.
• Patient-level metadata.
• Non-public reproduction packages.
• Internal training launch recipes or cluster-specific paths.
• Task-specific fine-tuned checkpoints unless separately released.

Citation

Please cite the final Nature Biomedical Engineering record once it is live:

@article{bonecot2026,
  title = {BoneCoT: Multi-center validation of a whole-body skeleton foundation model for bone metastases guided by clinician-derived chain of thought},
  author = {Zhao, Hui and Zhang, Ruipeng and Wang, Zhiyu and Gu, Yifeng and Xu, Shengyuan and Wang, Sheng and Li, Yuehua},
  journal = {Nature Biomedical Engineering},
  year = {2026},
  doi = {10.1038/s41551-026-01736-1}
}

BoneFM builds on DINOv2-style self-supervised vision-transformer code. Please also cite the relevant DINOv2 work when using inherited implementation components.

License

The public BoneFM release is made available under CC BY-NC 4.0 for non-commercial research use, subject to any applicable third-party code licenses in the accompanying implementation.