This repository provides the official PyTorch implementation of TinyUSFM — the first lightweight ultrasound foundation model designed for efficient and deployable medical AI.
TinyUSFM achieves foundation-level representation capability with only 5.5M parameters and 2.16 GFLOPs, serves as the official lightweight continuation of our last work Ultrasound Foundation Model(USFM), preserving its generalization ability while enabling efficient deployment across diverse ultrasound applications.
🏆 TinyUSFM achieved 1st Place in the MICCAI 2025 Intrapartum Ultrasound Grand Challenge (IUGC).
🧩 We release UniUS-Bench, the largest public ultrasound benchmark covering 15 organs.
⚙️ This repository also fixes and improves the segmentation settings of USFM.
Key Features
- 🧠 Feature–Gradient Driven Coreset Selection — Curates high-quality, diverse ultrasound samples for efficient distillation.
- 🔄 Domain-Separated Masked Image Modeling — Preserves spatial and frequency representations critical for ultrasound.
- ⚖️ Consistency-Driven Dynamic Distillation — Transfers reliable teacher knowledge with adaptive weighting.
- ⚡ Efficient Deployment — Matches USFM performance with only 6% of parameters and computation.
# Clone this repo
git clone https://github.com/MacDunno/TinyUSFM.git
cd TinyUSFM
# Create environment
conda create -n tinyusfm python=3.12
conda activate tinyusfm
# install pytorch according to instructions
# https://pytorch.org/get-started/
pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu126
# Install dependencies
pip install -r requirements.txtUniUS-Bench integrates 8 classification and 10 segmentation datasets (≈60,940 images, 15 organs) for standardized evaluation of ultrasound foundation models.
| Dataset | Organ(s) | #Images | Link |
|---|---|---|---|
| CUBS | Carotid artery | 1,378 | mendeley |
| UF1990 | Uterus | 1,990 | mendeley |
| TN3K | Thyroid | 3,491 | github |
| STMUS | Skeletal muscle | 5,312 | mendeley |
| AUL | Liver | 735 | zenodo |
| BUSI | Breast | 780 | homePage |
| MMOTU | Ovarian | 1,469 | github |
| Fetal Planes | Fetus | 12,400 | zenodo |
| Luminous | Multifidus muscle | 341 | homepage |
| KidneyUS | Kidney | 487 | github |
| GIST514 | Stomach | 514 | github |
| DDTI | Thyroid | 637 | github |
| BUSBRA | Breast | 1,875 | github |
| NerveSeg | Neck nerve | 5,735 | kaggle |
| LUSS | Lung | 564 | Leeds |
| FH-PS-AoP | Pelvis | 4,000 | zenodo |
| CAMUS | Cardiac | 19,232 | insa-lyon |
Pretrained TinyUSFM weights: Google Drive
Pretrained USFM weights: Google Drive
python train_cls.py --model_name TinyUSFM # or USFM python train_seg.py --model_name TinyUSFM_Seg # or USFM_SegThis project is licensed under the Apache 2.0 License.
See LICENSE for details.
If you find this work useful for your research, please cite:
@article{tinyusfm,
author={Ma, Chen and Jiao, Jing and Liang, Shuyu and Fu, Junhu and Wang, Qin and Li, Zeju and Wang, Yuanyuan and Guo, Yi},
journal={IEEE Journal of Biomedical and Health Informatics},
title={TinyUSFM: Towards Compact and Efficient Ultrasound Foundation Models},
year={2026},
pages={1-14},
doi={10.1109/JBHI.2026.3678309}
}
@article{usfm,
title={Usfm: A universal ultrasound foundation model generalized to tasks and organs towards label efficient image analysis},
author={Jiao, Jing and Zhou, Jin and Li, Xiaokang and Xia, Menghua and Huang, Yi and Huang, Lihong and Wang, Na and Zhang, Xiaofan and Zhou, Shichong and Wang, Yuanyuan and others},
journal={Medical image analysis},
volume={96},
pages={103202},
year={2024},
publisher={Elsevier}
}
@incollection{iugc,
title={Unlabeled Data-Driven Fetal Landmark Detection in Intrapartum Ultrasound},
author={Ma, Chen and Li, Yunshu and Guo, Bowen and Jiao, Jing and Huang, Yi and Wang, Yuanyuan and Guo, Yi},
booktitle={Intrapartum Ultrasound Grand Challenge},
pages={14--23},
year={2025},
publisher={Springer}
}For any questions, please feel free to contact: cma24@m.fudan.edu.cn