StyleMapper: Deep Learning for Breast MRI Style Transfer with Limited Training Data

This is the code for our paper Deep Learning for Breast MRI Style Transfer with Limited Training Data, Journal of Digital Imaging 2022.

Code built on https://github.com/NVlabs/MUNIT.

Citation

If you use this code, please cite our paper:

@article{cao2022deep,
  title={Deep Learning for Breast MRI Style Transfer with Limited Training Data},
  author={Cao, Shixing and Konz, Nicholas and Duncan, James and Mazurowski, Maciej A},
  journal={Journal of Digital Imaging},
  pages={1--13},
  year={2022},
  publisher={Springer}
}

File Descriptions

• commands.txt: example commands for training and experiments/testing
• requirements.txt: required packages that you (may) need to install
• train.py: script for training the model
• test.py: script for testing the model
• style_code.py: script for extracting style codes from images using trained model
• compare_stylecodes.py: script for comparing style codes of images, used for experiments in the supplementary material
• validation.py: script with validation utilities
• trainer.py: script with training utilities
• networks.py: script with model building utilities
• data.py: script for loading data
• transforms.py: script with image transformations used by model
• utils.py: script with miscellaneous utilities
• configs: Config files for training and testing with the same settings as in the paper. See commands.txt for examples of how to use these.

Data Setup Instructions

All data is from the Duke Breast Cancer MRI dataset, linked here. Once the data is downloaded, it can be sorted by MRI manufacturer (GE or Siemens) via the third column of https://wiki.cancerimagingarchive.net/download/attachments/70226903/Clinical_and_Other_Features.xlsx.

Next, you can create folders named datasets/breast_mri/GE and datasets/breast_mri/SIEMENS in the base directory to house post-contrast MRI slice DICOM files for scans from these two scanner manufacturers, downloaded via the previous links. You can divide these files into data subsets with folders in datasets/breast_mri/GE of trainA, trainB, testA, testB, validationA, validationB, where A and B are used to split a subset in half (to more easily load into the model), e.g. the training set. For example, an MRI slice image in the training set could be datasets/breast_mri/GE/trainA/Breast_MRI_438_post_1_100.dcm.