Adaptive Correspondence Scoring for Unsupervised Medical Image Registration

[Paper] [Project page] [Poster] [Slides]

Motivation

Framework

Todo

[Dec. 2024] - We are in the process of organizing our repository.

Installation

Please pip install any other missing packages.

conda create -f requirements.yml -n AdaCS_env # Create an environment named AdaCS_env
conda activate AdaCS_env

# configure wandb
# pip install wandb==0.13.10 

Dataset

Please download ACDC dataset [website] and CAMUS dataset [website] from their websites.

Default directory structure

├── Dataset                   
|   ├── ACDC       # Place the downloaded dataset here
|   |   ├── train
|   |   ├── val
|   |   ├── test
|   ├── CAMUS
|   |   ├── train
|   |   ├── ...
├── Code
|   ├── AdaCS
|   |   ├── train_vxm.py
|   |   ├── test_vxm.py
|   |   ├── train_tsm.py
|   |   ├── ...

Train

We currently support voxelmorph-based approach for review purpose and we will open-source other architectures upon acceptance.

python train_vxm.py --dataset YOUR_DATASET  --bidir --model-dir YOUR_MODEL_SAVE_DIR --motion-loss-type 'wmse' --scoring-loss-type 'scoringwmse'  --warm_start --wandb-name YOUR_PROJECT_NAME 

Test

python test_vxm.py --dataset YOUR_DATASET --test-dir '../../Dataset/ACDC/test/' --result-dir YOUR_RESULT_SAVE_DIR --model-motion YOUR_SAVED_MOTION_WEIGHT --model-scoring YOUR_SAVED_SCORING_WEIGHT --inshape 128 128

Citation

@inproceedings{zhang2024adaptive,
  title={Adaptive Correspondence Scoring for Unsupervised Medical Image Registration},
  author={Zhang, Xiaoran and Stendahl, John C and Staib, Lawrence H and Sinusas, Albert J and Wong, Alex and Duncan, James S},
  booktitle={European Conference on Computer Vision},
  pages={76--92},
  year={2024},
  organization={Springer}
}

Acknowledgement

We use implementation of Voxelmorph, Transmorph and Diffusemorph.