TransMatch

This repository contains the implementations of the paper 'TransMatch: A Transformer-based Multilevel Dual-Stream Feature Matching Network for Unsupervised Deformable Image Registration,' published in the IEEE Transactions on Medical Imaging (TMI) journal.

[Paper] [Code]

What's News

[2/25/2025] - [TODO] Cross-modal image registration supported will coming soon.

[12/17/2023] - The code has been re-reviewed and debugged, and it is now ready for direct training and inference on the provided LPBA40 example dataset.

Descriptions

This repository currently provides examples of implementing, training, and inferring with the core model code. It also includes guidance on running the code on sample datasets. Additionally, there are suggestions and code for visual analysis.

Train and Infer Command

Before running the commands, please ensure that the dataset has been correctly placed. Taking the example of running the sample code on the LPBA40 dataset, ensure that the LPBA40 dataset is placed under ../../../Dataset/LPBA40_delineation/. This will ensure that the code can run directly without encountering any path-related errors. (Here, ./ refers to the directory path where Train.py and Infer.py are located.)

For Linux:

Train:
export CUDA_VISIBLE_DEVICES=0 (If needed)
python Train.py

Infer:
export CUDA_VISIBLE_DEVICES=0 (If needed)
python Infer.py

Dataset

LPBA40 Datasets is uploded now. [LPBA40 datasets download link]

Additionally, you can effortlessly train the code by customizing your datasets. Please be aware that you'll need to adjust the code file located at /utils/datagenerators_atlas.py to ensure the dataset is loaded in accordance with your specific dataset organization format.

TODO

• Core implementation code
• Description of run script
• Visualization code
• Datasets url link (LPBA40 is uploaded)
• Cross-modal image registration supported
• Docker images & Google Colab Documents (Pull requests are always welcome!)

Contact

Feel free to contact me if you have any further questions: snowbplus [AT] gmail [DOT] com.

Citation

If you find this code is useful in your research, please consider to cite:

@article{chen2023transmatch,
  title={Transmatch: A transformer-based multilevel dual-stream feature matching network for unsupervised deformable image registration},
  author={Chen, Zeyuan and Zheng, Yuanjie and Gee, James C},
  journal={IEEE transactions on medical imaging},
  volume={43},
  number={1},
  pages={15--27},
  year={2023},
  publisher={IEEE}
}

Acknowledgements

Thanks to Junyu Chen and Tony C. W. Mok for their guidance and help with the open source code.