Reimplement of mm-gan.
The origin implement can be seen in https://github.com/trane293/mm-gan
Related article: Missing MRI Pulse Sequence Synthesis using Multi-Modal Generative Adversarial Network.
MM-GAN is a novel GAN-based approach that allows synthesizing missing pulse sequences (modalities) for an MRI scan. For more details, please refer to the paper.
To cite the paper (IEEE TMI):
@ARTICLE{Sharma20,
author={A. {Sharma} and G. {Hamarneh}},
journal={IEEE Transactions on Medical Imaging},
title={Missing MRI Pulse Sequence Synthesis Using Multi-Modal Generative Adversarial Network},
year={2020},
volume={39},
number={4},
pages={1170-1183},}
To cite the origin repository:
@misc{Sharma20Code,
author = {A. {Sharma}},
title = {MM-GAN: Missing MRI Pulse Sequence Synthesis using Multi-Modal Generative Adversarial Network},
year = {2020},
publisher = {GitHub},
journal = {GitHub Repository},
howpublished = {\url{https://github.com/trane293/mm-gan}}
}
In order to run the code, we recommend that you use Anaconda distribution. We provivde an environment.yml file that can be used to recreate the exact same Python environment that can be used to run this code.
Once Anaconda is installed, simply do:
conda env create -f environment.ymlDownload BRATS2018 dataset from the official website https://www.med.upenn.edu/sbia/brats2018/data.html and extract both the Training and Validation zip files to a folder, say:
data/BRATS2018/Training/
data/BRATS2018/Training/HGG
data/BRATS2018/Training/LGG
Validation patients do not have HGG/LGG labels.
data/BRATS2018/Validation/
Once done, switch to the conda virtualenv, and run:
python modules/create_hdf5_file.py
Once it's finished, run:
python preprocess.py
You should have two HDF5 datasets now, one of which is raw BRATS2018 dataset, and one that is cropped to coordinates.
In order to train MM-GAN on HGG data do:
./mmgan_hgg.shSimilarly for training on LGG data:
./mmgan_lgg.shAnd it should start the network training process.
| Folder Name | Purpose |
|---|---|
| modules | contains modules and helper functions, along with network architectures |
| modules/advanced_gans | contains the main Pix2Pix implementation |
| notebooks | contains rough notebooks for quick prototyping |
| prep_BRATS2015 | contains code to prepare BRATS2015 data for training/evaluation |
| prep_ISLES2015 | contains code to prepare ISLES2015 data for training/validation |