How to get it working

Here's a guide how to evaluate the provided SIFA model, which was trained on MRI images and adapted to the domain of CT images. We test how well the adapted model performs on provided CT images.

Follow along!

Step 1/4: Download the testing data

Go to Google Drive and download the .zip archive. This might take 1 minute. Extract the archive into the folder

data/test_ct_image&labels/

This will yield the following 8 files:

data/test_ct_image&labels/gth_ct_1003.nii.gz
data/test_ct_image&labels/gth_ct_1008.nii.gz
data/test_ct_image&labels/gth_ct_1014.nii.gz
data/test_ct_image&labels/gth_ct_1019.nii.gz
data/test_ct_image&labels/image_ct_1003.nii.gz
data/test_ct_image&labels/image_ct_1008.nii.gz
data/test_ct_image&labels/image_ct_1014.nii.gz
data/test_ct_image&labels/image_ct_1019.nii.gz

Step 2/4: Download the model weights

Also, go to Dropbox and scroll to the bottom. Download the following files:

sifa-cardiac-mr2ct.data-00000-of-00001
sifa-cardiac-mr2ct.index
sifa-cardiac-mr2ct.meta

This might take 2 minutes.

Place these three files into the local directory SIFA-model.

Step 3/4: Build a Docker image.

Run

./INSTALL

This will create a new Docker image in your local Docker installation.

Step 4/4: Run the Docker image.

Run

./RUN

The command might not stop for about 20 minutes. You might need to be patient. It will write 4 files to data/test_ct_image&labels. Eventually, you might get an output like the following:

Dice:
AA :78.3(3.0)
LAC:77.5(5.3)
LVC:73.1(8.6)
Myo:61.2(7.9)
Mean:72.5
ASSD:
AA :9.3(1.6)
LAC:8.7(3.7)
LVC:7.0(2.4)
Myo:6.9(2.1)
Mean:8.0

---

Unsupervised Bidirectional Cross-Modality Adaptation via Deeply Synergistic Image and Feature Alignment for Medical Image Segmentation

Tensorflow implementation of our unsupervised cross-modality domain adaptation framework.
This is the version of our TMI paper.
Please refer to the branch SIFA-v1 for the version of our AAAI paper.

Paper

Unsupervised Bidirectional Cross-Modality Adaptation via Deeply Synergistic Image and Feature Alignment for Medical Image Segmentation
IEEE Transactions on Medical Imaging

Installation

• Install TensorFlow 1.10 and CUDA 9.0
• Clone this repo

git clone https://github.com/cchen-cc/SIFA
cd SIFA

Data Preparation

• Raw data needs to be written into tfrecord format to be decoded by ./data_loader.py. The pre-processed data has been released from our work PnP-AdaNet. The training data can be downloaded here. The testing CT data can be downloaded here. The testing MR data can be downloaded here.
• Put tfrecord data of two domains into corresponding folders under ./data accordingly.
• Run ./create_datalist.py to generate the datalists containing the path of each data.

Train

• Modify the data statistics in data_loader.py according to the specifc dataset in use. Note that this is a very important step to correctly convert the data range to [-1, 1] for the network inputs and ensure the performance.
• Modify paramter values in ./config_param.json
• Run ./main.py to start the training process

Evaluate

• Our trained models can be downloaded from Dropbox. Note that the data statistics in evaluate.py need to be changed accordingly as specificed in the script.
• Specify the model path and test file path in ./evaluate.py
• Run ./evaluate.py to start the evaluation.

Citation

If you find the code useful for your research, please cite our paper.

@article{chen2020unsupervised,
  title     = {Unsupervised Bidirectional Cross-Modality Adaptation via 
               Deeply Synergistic Image and Feature Alignment for Medical Image Segmentation},
  author    = {Chen, Cheng and Dou, Qi and Chen, Hao and Qin, Jing and Heng, Pheng Ann},
  journal   = {arXiv preprint arXiv:2002.02255},
  year      = {2020}
}

@inproceedings{chen2019synergistic,
  author    = {Chen, Cheng and Dou, Qi and Chen, Hao and Qin, Jing and Heng, Pheng-Ann},
  title     = {Synergistic Image and Feature Adaptation: 
               Towards Cross-Modality Domain Adaptation for Medical Image Segmentation},
  booktitle = {Proceedings of The Thirty-Third Conference on Artificial Intelligence (AAAI)},
  pages     = {865--872},
  year      = {2019},
}

Acknowledgement

Part of the code is revised from the Tensorflow implementation of CycleGAN.

Note

• The repository is being updated
• Contact: Cheng Chen ([email protected])