Access CIRCLE
CIRCLE is a research model for chest CT image understanding. Please review and acknowledge the license, responsible-use terms, and model limitations before accessing the model.
CIRCLE Model Access Terms and Responsible Use Notice
CIRCLE is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) license.
Access to the official CIRCLE model repository is gated to promote responsible use and to inform users of the model's intended scope, limitations, and responsibilities.
- License and Non-Commercial Use
You agree to comply with the CC BY-NC-SA 4.0 license, including applicable attribution, non-commercial use, and ShareAlike requirements. Commercial use requires separate permission from the model providers.
- Intended Use
CIRCLE is developed and released primarily for scientific research, academic study, education, benchmarking, and other non-commercial research purposes.
- Medical and Clinical Disclaimer
CIRCLE is a research model and has not been validated, certified, cleared, or approved as a medical device.
The model and its outputs are not intended to replace the judgment of qualified healthcare professionals and should not be relied upon as the sole basis for diagnosis, treatment, triage, prognosis, patient management, or other clinical decision-making.
- Model Limitations
CIRCLE may generate inaccurate, incomplete, inconsistent, biased, hallucinated, or otherwise erroneous outputs. Performance may vary across patient populations, institutions, scanners, imaging protocols, and downstream applications.
Users are responsible for independently validating the model for their intended use.
- Privacy, Ethics, and Medical Data
Users are responsible for complying with applicable laws, regulations, privacy requirements, research ethics, institutional policies, and approvals when using CIRCLE, particularly when processing medical or patient-related data.
Users should not process personally identifiable or protected health information without appropriate authorization and safeguards.
- Re-identification and Privacy Protection
Users must not intentionally use CIRCLE to identify or re-identify individuals or to extract, reconstruct, or infer personally identifiable or protected health information from the model or its outputs.
- Attribution
Users must appropriately acknowledge CIRCLE and cite the associated CIRCLE publication in scholarly and public research outputs.
- Disclaimer and Responsibility
CIRCLE is provided "as is" and "as available", without warranties of accuracy, completeness, reliability, fitness for a particular purpose, or clinical performance, to the extent permitted by applicable law.
Users remain responsible for evaluating whether CIRCLE is appropriate for their intended use and for the consequences of using the model or its outputs.
Log in or Sign Up to review the conditions and access this model content.
🤗 CIRCLE: A Comprehensive Foundation Model for Automated Chest CT Interpretation
🔥 CIRCLE Dataset | 💻 GitHub Code | 📄 Paper is coming soon!
Overview
CIRCLE is a thoracic CT foundation model developed through a collaboration among 15 Chinese hospitals and United Imaging Intelligence (UII). It combines image–text contrastive learning with supervised learning over 37 thoracic abnormality categories and is designed for non-contrast chest CT scans.
This repository now provides both Chinese and English CIRCLE models, together with corresponding Chinese and English CIRCLE-Labeler models.
Research use only. These models are not medical devices and must not be used as a substitute for professional diagnosis or clinical decision-making.
Released Models
| Component | Directory | Input language | Backbone / contents |
|---|---|---|---|
| CIRCLE Chinese | models/CIRCLE-zh |
Chinese reports and prompts | 3D vision encoder + Chinese RoBERTa text encoder |
| CIRCLE English | models/CIRCLE-en |
English reports and prompts | 3D vision encoder + RadBERT-RoBERTa-4m text encoder |
| CIRCLE-Labeler Chinese | models/CIRCLE-labeler-zh |
Chinese radiology reports | Qwen3-1.7B-based report labeler |
| CIRCLE-Labeler English | models/CIRCLE-labeler-en |
English radiology reports | Qwen3-1.7B-based report labeler |
Repository layout:
models/
├── CIRCLE-zh/
│ ├── vision_encoder.bin
│ └── text_encoder/
├── CIRCLE-en/
│ ├── vision_encoder.bin
│ └── text_encoder/
├── CIRCLE-labeler-zh/
└── CIRCLE-labeler-en/
The English CIRCLE model was trained with English reports translated from the Chinese training reports using DeepSeek-V4 Flash. The image-side architecture and preprocessing are the same for both language variants. For text-dependent tasks, always pair each vision encoder with its corresponding language-specific text encoder and use reports or prompts in the same language.
Both CIRCLE-Labeler variants were obtained by fine-tuning Qwen3-1.7B on approximately 400,000 chest CT reports.
How to Use
For detailed instructions on how to use this project, please refer to the official GitHub repository:
- GitHub Repository: https://github.com/uii-open-source/CIRCLE
The repository includes:
- Setup and installation guides
- Usage examples
- Configuration options
Make sure to check the README.md for the most up-to-date information.
Evaluation Highlights
Validation performance for 37-abnormality detection:
| Cohort | Mean AUC |
|---|---|
| Internal cohort (50,654 volumes) | 0.916 |
| External Set I (61,223 volumes) | 0.882 |
| External Set II (31,156 volumes) | 0.875 |
| External Set III (1,099 volumes) | 0.872 |
Additional evaluated capabilities include:
- zero-shot diagnosis, including prompt-based and KNN-based settings;
- patient-level screening;
- cross-modal CT ↔ report retrieval;
- CIRCLE-Report and CIRCLE-VQA transfer;
- segmentation transfer through nnUNet adaptation.
For complete protocols and evaluation scripts, refer to the official code repository.
License
All released CIRCLE components are licensed under CC BY-NC-SA 4.0.
You may share and adapt the models for non-commercial research provided that you:
- give appropriate credit and indicate any changes;
- distribute derivatives under the same license;
- do not use the models or datasets commercially without explicit permission.
Citation
If you use CIRCLE in your research, please cite the upcoming paper and acknowledge the project:
@misc{circle2025,
author = {{United Imaging Intelligence (UII)} and {Zhongshan Hospital, Fudan University}},
title = {{CIRCLE}: A Comprehensive Foundation Model for Automated Chest CT Interpretation},
year = {2025},
note = {Paper under submission}
}
If this work is useful to you, please consider starring the GitHub repository. ⭐