📖 TL;DR

This work introduces VasoMIM, a vascular anatomy-aware self-supervised learning framework designed specifically for X-ray angiogram pre-training. To support this, we curated XA-170K, the largest existing X-ray angiogram dataset. VasoMIM is validated on four downstream tasks crucial for X-ray angiogram analysis, demonstrating superior performance.

✨ News

• 2026-03 🚀 XA-170K is now available on Hugging Face.
• 2026-02 💻 We released the code on GitHub.
• 2026-02 📝 We posted the journal version of VasoMIM on arXiv.
• 2025-11 🎉 VasoMIM-v1 was accepted to AAAI 2026.

🛠️ Method

⚙️ Requirements

• This repository is a modification of the official MAE repository. Installation and environment preparation steps follow the original repo.
• Note on Timm: This code relies on timm==1.0.20.

💾 Datasets

1. Pre-training Dataset (XA-170K)

XA-170K aggregates data from four publicly available sources: CADICA, SYNTAX, XCAD, and CoronaryDominance.

Option A: Direct Download (Recommended) You can download the curated XA-170K dataset directly from our Hugging Face repo.

Option B: Manual Collection Alternatively, you can collect the raw data from the original sources:

Dataset	Images	Link
CADICA	6,594	Download
SYNTAX	2,943	Download
XCAD	1,621	Download
CoronaryDominance	160,320	Download
Total	171,478	-

Directory Structure The XA-170K dataset should be organized as follows:

/path/to/XA-170K/
  ├── cadica/
  │    ├── image1.png
  │    └── ...
  ├── cadica_frangi/
  │    ├── image1.png
  │    └── ...
  ├── syntax/
  ├── syntax_frangi/
  ├── xcad/
  ├── xcad_frangi/
  ├── coronarydominance/
  └── coronarydominance_frangi/

2. Downstream Datasets

Dataset	Train	Test	Link	Task
ARCADE-V	1,000	3,00	Download	Vessel Segmentation
CAXF	337	201	In-house*	Vessel Segmentation
XCAV	175	46	Download	Vessel Segmentation
ARCADE-S	1,000	3,00	Download	Stenosis Segmentation
ARCADE-VS	1,000	3,00	Download	Vessel Segment Segmentation
Stenosis	7,492	833	Download	Stenosis Detection

*Note: Please contact De-Xing Huang (huangdexing2022@ia.ac.cn) if you wish to use CAXF for research purposes.

🚀 Pre-training

We pre-trained VasoMIM on 8 x NVIDIA H20 GPUs (96 GB).

cd /path/to/this/workspace
./pretrain_vasomim.sh

📦 Pre-trained Models

Coming Soon!

🙌 Acknowledgement

This project is built upon MAE and HPM. For segmentation task, our implementations are based on CheXWorld. For detection task, we utilize ViTDet from Detectron2. The Frangi filter implementation is adapted from DeNVeR.

We thank the authors of these repositories for their wonderful work.

✏️ Citation

If you find VasoMIM useful for your research, please consider citing our paper:

@inproceedings{huang2026vasomim,
  title={{VasoMIM}: Vascular anatomy-aware masked image modeling for vessel segmentation},
  author={Huang, De-Xing and others},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2026}
}

@article{huang2026vascular,
  title={Vascular anatomy-aware self-supervised pre-training for X-ray angiogram analysis},
  author={Huang, De-Xing and others},
  journal={arXiv preprint arXiv:2602.11536},
  year={2026}
}

If you utilize the pre-training dataset, please also consider citing the original data sources:

@article{jimenez2024cadica,
  title={CADICA: A new dataset for coronary artery disease detection by using invasive coronary angiography},
  author={Jim{\'e}nez-Partinen and others},
  journal={Expert Systems},
  volume={41},
  number={12},
  pages={e13708},
  year={2024}
}

@article{mahmoudi2025x,
  title={X-ray Coronary Angiogram images and {SYNTAX} score to develop Machine-Learning algorithms for {CHD} Diagnosis},
  author={Mahmoudi, Seyed Sajjad and others},
  journal={Scientific Data},
  volume={12},
  number={1},
  pages={471},
  year={2025}
}

@inproceedings{ma2021self,
  title={Self-supervised vessel segmentation via adversarial learning},
  author={Ma, Yuxin and others},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  pages={7536--7545},
  year={2021}
}

@article{kruzhilov2025coronarydominance,
  title={{CoronaryDominance}: Angiogram dataset for coronary dominance classification},
  author={Kruzhilov, Ivan and others},
  journal={Scientific Data},
  volume={12},
  number={1},
  pages={341},
  year={2025}
}