Unpaired abdomen CT registration

Note: Please read the DeepReg Demo Disclaimer.

Source Code

Author

DeepReg Development Team (Ester Bonmati)

Application

This demo shows how to register unpaired abdominal CT data from different patients using DeepReg. In addition, the demo demonstrates the difference between the unsupervised, weakly-supervised and their combination, using a U-Net.

Data

The data set is from the MICCAI Learn2Reg grand challenge (https://learn2reg.grand-challenge.org/) task 3 [1], and can be downloaded directly from https://learn2reg.grand-challenge.org/Datasets/.

Instruction

  • Install DeepReg;

  • Change the working directory to the root directory of DeepReg project;

  • Run [demo_data.py] to download and extract all files, and to split the data into training, validation and testing. If the data has already been downloaded. This will also download the pre-trained models:

python ./demos/unpaired_ct_abdomen/demo_data.py

After running the command you will have the following directories in DeepReg/demos/unpaired_ct_abdomen/dataset:

pre-trained  test  train  val

  • The next step is to train the network using DeepReg. To train the network, run one of the following commands in command line:

  • Unsupervised learning

deepreg_train --gpu "0" --config_path demos/unpaired_ct_abdomen/unpaired_ct_abdomen_unsup.yaml --log_dir unpaired_ct_abdomen_unsup

  • Weakly-supervised learning

deepreg_train --gpu "1" --config_path demos/unpaired_ct_abdomen/unpaired_ct_abdomen_weakly.yaml --log_dir unpaired_ct_abdomen_weakly

  • Combined learning

deepreg_train --gpu "2" --config_path demos/unpaired_ct_abdomen/unpaired_ct_abdomen_comb.yaml --log_dir unpaired_ct_abdomen_comb

  • After training the network, run demo_predict:

The following example uses a pre-trained model, on CPU.

deepreg_predict --gpu "" --config_path demos/unpaired_ct_abdomen/unpaired_ct_abdomen_unsup.yaml --ckpt_path demos/unpaired_ct_abdomen/dataset/pre-trained/unsup/weights-epoch5000.ckpt --log_dir unpaired_ct_abdomen_unsup --save_png --mode test

deepreg_predict --gpu "" --config_path demos/unpaired_ct_abdomen/unpaired_ct_abdomen_weakly.yaml --ckpt_path demos/unpaired_ct_abdomen/dataset/pre-trained/weakly/weights-epoch2250.ckpt --log_dir unpaired_ct_abdomen_weakly --save_png --mode test

deepreg_predict --gpu "" --config_path demos/unpaired_ct_abdomen/unpaired_ct_abdomen_comb.yaml --ckpt_path demos/unpaired_ct_abdomen/dataset/pre-trained/comb/weights-epoch2000.ckpt --log_dir unpaired_ct_abdomen_comb --save_png --mode test

  • Finally, prediction results can be seen in the respective test folders specified in deepreg_predict.

Pre-trained model

Three pre-trained models are available for this demo, for different training strategies described above. These will be downloaded in respective sub-folders under the /dataset folder using the demo_data.py. Run the user-trained model by specifying with --ckpt_path the location where the ckpt files will be saved, in this case (specified by deepreg_train as above), /logs/unpaired_ct_abdomen_unsup/, /logs/unpaired_ct_abdomen_weakly/ or /logs/unpaired_ct_abdomen_comb/.

Tested DeepReg version

Last commit at which demo was tested: 3157f880eb99ce10fc3a4a8ebcc595bd67be24e1

Contact

Please raise an issue.

Reference

[1] Adrian Dalca, Yipeng Hu, Tom Vercauteren, Mattias Heinrich, Lasse Hansen, Marc Modat, Bob de Vos, Yiming Xiao, Hassan Rivaz, Matthieu Chabanas, Ingerid Reinertsen, Bennett Landman, Jorge Cardoso, Bram van Ginneken, Alessa Hering, and Keelin Murphy. (2020, March 19). Learn2Reg - The Challenge. Zenodo. http://doi.org/10.5281/zenodo.3715652