Deep learning network-based registration method applied on cardiac motion tracking from cardiac MR images (cMRI). If you use this code or part of this code, please consider citing the following papers:

Qiu, H., Qin, C., Le Folgoc, L., Hou, B., Schlemper, J., Rueckert, D.:
Deep Learning for Cardiac Motion Estimation: Supervised vs. Unsupervised Training
STACOM Workshop, MICCAI 2019.
(STACOM19 version of the code can be found in branch stacom19)

Qin, C., Bai, W., Schlemper, J., Petersen, S.E., Piechnik, S.K., Neubauer, S., Rueckert, D.:
Joint learning of motion estimation and segmentation for cardiac MR image sequences
MICCAI 2018

Code developed and tested on Ubuntu 16.04 & 18.04 operating systems, using Python 3.6 and Pytorch 1.0.

To install the Python dependencies, run the following in the root directory of the repo after cloning the repo:

pip3 install -r requirements.txt

CUDA and cuDNN are required (tested with CUDA 9.0.176 and cuDNN 7.1.4). The code should work with any CUDA and cuDNN versions supported by Pytorch 1.0. Please refer to Pytorch and NVIDIA websites.

The code works on a model-directory-basis. Training, testing and inference of a model are all based on the model directory of this model. Logs, trained models, testing and inference results are all saved in the model directory.

Training:

python cardiac_motion/train.py --gpu [gpu_num] --model_dir [path_to_model_dir]

Testing (on the end-diastolic and end-systolic frames):

python cardiac_motion/eval.py --gpu [gpu_num] --model_dir [path_to_model_dir] --restore_file [file_name_of_saved_model]

Inference (on all frames of the sequences):

python cardiac_motion/inference.py --gpu [gpu_num] --model_dir [path_to_model_dir] --data_dir [path_to_data_dir]

Most setting parameters related to data or model are specified in the params.json file, which should be supplied in the model directory. This file is parsed into attributes of the object params in the code to pass the parameters. An example of this file is provided in the repo root directory.

Models trained on cardiac MR image data from the UK Biobank Imaging Study is available. Please feel free to email us to enquire if you are interested.

If you have any question regarding the paper or the code, feel free to open an issue in this repo or email us at: [email protected]