Skip to content

Weakly-supervised Biomechanically-constrained CT/MRI Registration of the Spine

License

Notifications You must be signed in to change notification settings

BailiangJ/spine-ct-mr-registration

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Weakly-supervised Biomechanically-constrained CT/MRI Registration of the Spine

This is the official Pytorch implementation of my paper: "Weakly-supervised Biomechanically-constrained CT/MRI Registration of the Spine" (MICCAI 2022)


Guidelines

  1. Check Registry and Config of MMEngine.
  2. Check the Config files for dataset, network, training and inference.
  3. Construct the data files following the structure shown below. Change the parameters in the dataset Config.
  4. Train the model with only image similarity and smoothness regularization (e.g. 400 epochs).
  5. Train the model with additionally rigidity losses to finetune. Check the supplementary materiel for weights of different rigidity losses.
  6. Inference on test data.

Prerequisites

  • Python 3.8.8
  • Numpy 1.23.4
  • Scipy 1.9.3
  • Pytorch 1.13.0
  • MONAI 1.0.1
  • MMCV 1.7.0
  • MMENGINE 0.5.0
  • WANDB 0.13.10
  • SimpleITK 2.0.2

Data structure

├──data_dir
│   ├──train
│   │   ├──001
│   │   │   ├──ct.nii.gz
│   │   │   ├──ct_mask.nii.gz
│   │   │   ├──mr_t1.nii.gz
│   │   │   ├──mr_mask.nii.gz
│   │   ├──002
...
│   ├──val
│   │   ├──...
...
│   ├──test
│   │   ├──...

Dataset

Since the dataset used in the paper is confidential and belongs to the hospital, we cannot share it. If you are also doing spine registration and trying to construct your own dataset, it's recommended to use the free spine CT vertebra segmentation webtool Anduin.


Pretrained Model

The pretrained model weights are also unable to share due to the confidentiality of the dataset.


Logging

I use Weight&Biases for logging in training. It's convenient to use (view the plot without ssh to port) and free. If you prefer to use tensorboard, you could also construct your own logger and add it to run_epoch.


Baseline method

Check Elastix.


Citation

If you find this repository useful in your research, please consider to cite use in your work by:

@inproceedings{jian2022weakly,
  title={Weakly-supervised Biomechanically-constrained CT/MRI Registration of the Spine},
  author={Jian, Bailiang and Azampour, Mohammad Farid and De Benetti, Francesca and Oberreuter, Johannes and Bukas, Christina and Gersing, Alexandra S and Foreman, Sarah C and Dietrich, Anna-Sophia and Rischewski, Jon and Kirschke, Jan S and others},
  booktitle={Medical Image Computing and Computer Assisted Intervention--MICCAI 2022: 25th International Conference, Singapore, September 18--22, 2022, Proceedings, Part VI},
  pages={227--236},
  year={2022},
  organization={Springer}
}

Acknowledgement

Many thanks to the following repositories for providing helpful resources to my work:


License & Copyright

© Bailiang Jian Licensed under the MIT Licensce

About

Weakly-supervised Biomechanically-constrained CT/MRI Registration of the Spine

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages