Diffeomorphic demons in TensorFlow

This is a course project in Medical Image Analysis on diffeomorphic demons and is implemented in TensorFlow. The lecture note for illustration and presentation is also included.

Getting Started

Project structure

The project contains 2D and 3D versions of Diffeomorphic demons and a dedicated version for circle-to-C registration.

Diffeomorphic-demons
|-- src      
|    |-- circle2C_reg.py                   # circle-to-C registration
|    |-- compare_metrics.py                # compare metrics from different settings
|    |-- crop_ROI.py                       # image preprocessing
|    |-- trainer_2d.py                     # 2D version of Diffeomorphic-demons
|    |-- trainer_3d.py                     # 3D version of Diffeomorphic-demons
|    |-- transformer.py                    # spatial transformer modules adapted from VoxelMorph
|    |-- utils.py                          # utility functions for preprocessing and evaluation

Usage

Registering a 3D fixed image with a 3D moving image is achieved by:

python trainer_3d.py
--demons_type diffeomorphic               # the transformation model for demons
--demons_force symmetric                  # demons force type
--regularizer fluid                       # regularization type for the displacement field
--normalization z-score                   # image normalization type during preprocessing
--max_length 2.                           # the maximum step length
--exp_steps 8                             # the number of exponential steps for the vector field
--training_iters 30                    
--display_steps 5
--cuda_device -1                          # use CPU only

Acknowledgement

Some parts of the code were adapted from VoxelMorph, which is for deep-learning-based medical image registration.

Citation

If you found the repository useful, please cite the lecture note as below:

@misc{Luo2019DiffeomorphicDemons,
  title={Medical Image Analysis, Diffeomorphic Demons},
  author={Xinzhe Luo},
  year={2019}
}

Contact

For any questions or problems please open an issue on GitHub.