Biomechanical Constraints Assimilation in Deep-Learning Image Registration

Supporting code for reproducing Biomechanical Constraints Assimilation in Deep-Learning Image Registration: Application to sliding and locally rigid deformations

Install dependencies

• conda create --name biomechanical_DLIR python=3.10
• conda activate biomechanical_DLIR
• pip install -r requirements.txt

Download Learn2Reg AbdomenCTCT data and preprocess (skip to Training Script)

Part of our results were obtained on the Learn2Reg AbdomenCTCT dataset.

1. Download data : (or manually here )

• mkdir Datasets/
• wget -P Datasets/ https://cloud.imi.uni-luebeck.de/s/32WaSRaTnFk2JeT/download/AbdomenCTCT.zip or curl ...
• unzip Datasets/AbdomenCTCT.zip -d Datasets/
• rm -r Datasets/AbdomenCTCT.zip

2. Generate regularization masks indicating which loss (rigidity,strain,Jacobian) to use for every voxel, as well as the sliding motion directions:

• python AbdomenCTCT_data_generation/reorient_images.py
• python AbdomenCTCT_data_generation/segmentTotalSegmentator.py -> Note that GPU usage is highly recommended..
• python AbdomenCTCT_data_generation/make_physics_loss_masks.py

You should now have the complete dataset under Datasets/AbdomenCTCT_reoriented/ including a /segmentationsTr folder.

Training and Evaluation Script

Model architectures, training orchestration and dataset configurations are handled through a json file. The main training script then executes the desired training. We provide all the configurations used for our hyper-parameter grid-search, specifically:

• configurations -> scripts for synthetic rotation dataset.
• configurations -> scripts for synthetic shearing dataset.
• configurations -> scripts for AbdomenCTCT dataset.

Model training is then simply achieved by : python biomechanical_DLIR/main.py -m biomechanical_DLIR/configs/simulated_configs_grid/RigidityDet/MSE_strainDet0_99.json

Models and logs will be saved under a new directory named results/ .

Finally, models can be evaluated with eval.py (or for full resolution on the AbdomenCTCT dataset: eval_full_res.py).

For more information: -> python biomechanical_DLIR/main.py -h -> python biomechanical_DLIR/eval.py -h -> python biomechanical_DLIR/eval_full_res.py -h

Custom Dataset or Models

Implementing a custom dataset requires adding a custom dataset class , then calling that dataset in the json object. Likewise, custom models can be added.

📝 Citation

@misc{kheil2025biomechanicalconstraintsassimilationdeeplearning,
      title={Biomechanical Constraints Assimilation in Deep-Learning Image Registration: Application to sliding and locally rigid deformations}, 
      author={Ziad Kheil and Soleakhena Ken and Laurent Risser},
      year={2025},
      eprint={2504.05444},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2504.05444}, 
}