Home

Awesome

<p align="center"><img width="90%" src="fig/title.jpg" /></p>

Medical image registration is a typical two-image task which requires specialized feature representation networks for deep-learning-based methods (The existing methods and their limitations have been evaluated in our papers). Therefore, we designed a X-shape feature representation backbone which combines the relationship-aware capacity of Transformer and the traits of two-image tasks which foucus not only on structure information of each image but also on cross correspondence between the image pair. The overall structure of our network is following:

<p align="center"><img width="90%" src="fig/XMorpher.jpg" /></p>

Paper

This repository provides the official implementation of XMorpher and its application under two different strategies in the following paper: <b>XMorpher: Full Transformer for Deformable Medical Image Registration via Cross Attention</b> <br/> Jiacheng Shi<sup>1</sup>, Yuting He<sup>1</sup>, Youyong Kong<sup>1,2,3</sup>, <br/> Jean-Louis Coatrieux<sup>1,2,3</sup>, Huazhong Shu<sup>1,2,3</sup>, Guanyu Yang<sup>1,2,3</sup>, and Shuo Li<sup>4</sup> <br/> <sup>1 </sup>LIST, Key Laboratory of Computer Network and Information Integration (Southeast University), Ministry of Education, Nanjing, China <br/> <sup>2 </sup>Jiangsu Provincial Joint International Research Laboratory of Medical Information Processing <br/> <sup>3 </sup>Centre de Recherche en Information Biomédicale Sino-Français (CRIBs) <br/> <sup>4 </sup>Dept. of Medical Biophysics, University of Western Ontario, London, ON, Canada <br/>

<b>International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI), 2022</b> <br/> paper | code | poster | video

Citation

If you use this code or use our pre-trained weights for your research, please cite our papers:

@inproceedings{shi2022xmorpher,
  title={XMorpher: Full Transformer for Deformable Medical Image Registration via Cross Attention},
  author={Shi, Jiacheng and He, Yuting and Kong, Youyong and Coatrieux, Jean-Louis and Shu, Huazhong and Yang, Guanyu and Li, Shuo},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={217--226},
  year={2022},
  organization={Springer}
}

Available implementation

★ Notes: implemented under two training strategies VoxelMorph and PC-Reg and the detailed corresponding main functions are Unsup-train.py and Semi-train.py respectively (Pytorch)

Major results from our work

  1. XMorpher has the best DSC score and Jacobian score under both strategies
<p align="center"><img width="90%" src="fig/table.png" /></p>
  1. XMorpher has visual superiority on some detailed structures
<p align="center"><img width="90%" src="fig/result.jpg" /></p>

Acknowledgement

This work was supported in part by the National Natural Science Foundation under grants (62171125, 61828101), CAAI-Huawei MindSpore Open Fund, CANN(Compute Architecture for Neural Networks), Ascend AI Processor, and Big Data Computing Center of Southeast University.