All-in-One Medical Image Re-Identification (CVPR2025)

January 11, 2026 · View on GitHub

MaMI is the fist unified re-identification model for medical images, capable of handling various imaging modalities such as X-ray, CT, fundus, and pathology images. By leveraging a Continuous Modality-based Parameter Adapter (ComPA) and integrating medical priors, MaMI supports both historical data-assisted diagnosis and privacy protection applications.

MAMI

Features

  • Unified Multi-Modality Model
    A single model supports multiple medical image modalities without having to train separate models for each modality.

  • Continuous Modality Adaptive Parameterization
    The ComPA module generates continuous modality representations to dynamically adapt model parameters based on the input image.

  • Medical Priors Integration
    Incorporates pre-trained Medical Foundation Models (MFMs) to enhance feature discrimination, capturing subtle identity-related cues for more robust re-identification.

Requirements

  • Python 3.7+
  • PyTorch 1.9+
  • CUDA-enabled GPU (recommended: NVIDIA RTX 4090)

Pre-trained Model

MAMI_pretrained.pth

Google Driver Link: https://drive.google.com/file/d/159KsDSgzgFCdSLN-I1iO0I2Z8Dm7dpiK/view?usp=sharing

BaiduYun (百度网盘) Link: https://pan.baidu.com/s/1GD-TsafqYhbXM6VwQVUwIA?pwd=fbge

Usage

  1. You can download the pre-trained model from the above link.

  2. copy the file the file "MAMI_pretrained.pth" to this directory.

  3. run "test.bash"

ToDo

  • Open Source Train/Validation Split
    Publish the scripts/configuration used to generate the train/val split for the dataset.

  • Open Source The Benchmark Methods

  • Open Source Training Code
    Release the complete training pipeline including all scripts and configuration files.

Citation

If you find our work useful, please consider citing our works on medical image privacy protection:

@inproceedings{tian2025towards,
  title={Towards All-in-One Medical Image Re-Identification},
  author={Tian, Yuan and Ji, Kaiyuan and Zhang, Rongzhao and Jiang, Yankai and Li, Chunyi and Wang, Xiaosong and Zhai, Guangtao},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={30774--30786},
  year={2025}
}


@inproceedings{tian2025semantic,
  title={Semantic versus Identity: A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification},
  author={Tian, Yuan and Wang, Shuo and Zhang, Rongzhao and Chen, Zijian and Jiang, Yankai and Li, Chunyi and Zhu, Xiangyang and Yan, Fang and Hu, Qiang and Wang, XiaoSong and others},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={20613--20625},
  year={2025}
}

@article{tian2025rofi,
  title={ROFI: a deep learning-based ophthalmic sign-preserving and reversible patient face anonymizer},
  author={Tian, Yuan and Zhou, Min and Chen, Yitong and Li, Fang and Qi, Lingzi and Wang, Shuo and Xu, Xieyang and Yu, Yu and Xu, Shiqiong and Lei, Chaoyu and others},
  journal={npj Digital Medicine},
  volume={8},
  number={1},
  pages={705},
  year={2025},
  publisher={Nature Publishing Group UK London}
}

@inproceedings{tian2025medical,
  title={Medical manifestation-aware de-identification},
  author={Tian, Yuan and Wang, Shuo and Zhai, Guangtao},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39},
  number={25},
  pages={26363--26372},
  year={2025}
}