reg-cyclical-self-train

Source code for our Miccai2023 paper Unsupervised 3D registration through optimization-guided cyclical self-training [pdf].

Dependencies

Please first install the following dependencies

• Python3 (we use 3.9.7)
• pytorch (we use 1.10.2)
• numpy
• scipy
• nibabel

Data Preparation

1. Download the Abdomen CT-CT dataset of the Learn2Reg challenge.
2. Modify the variable path in line 8 of data_utils.py such that it points to the root directory of the data.

Training

Execute python main.py --phase train --out_dir PATH/TO/OUTDIR --gpu GPU --num_warps 2 --ice true --reg_fac 1. --augment true --adam true --sampling true.

Testing

In l. 9 of test.py, set the path to the model weights you want to use for testing (for example our final_model.pth). Subsequently, execute python main.py --phase test --gpu GPU

Citation

If you find our code useful for your work, please cite the following paper

@article{unsupervised,
  title={Unsupervised 3D registration through optimization-guided cyclical self-training},
  author={Bigalke, Alexander and Hansen, Lasse and Mok, Tony C. W. and Heinrich, Mattias P},
  journal={arXiv preprint arXiv:2306.16997},
  year={2023}
}