Towards sharing brain images: Differentially private TOF-MRA images with segmentation labels using generative adversarial networks

Aim

The aim of the project is to generate shareable TOF-MRA image patches and corresponding segmentation labels with differential privacy guarantees.

Files

• config.py: configuration file for both GAN and U-net
• run_gan.py: script for running the GAN with differential privacy
• generate_samples.py: script synthesizing image-label pairs from GAN model
• model.py: GAN model file
• run_unet.py: script for running one or more U-nets based on (Livne et al.)
• eval_unet.py: script evaluating all specified U-nets on the validation set and the U-net with best validation performance on the test set. All results are printed into a .txt-file.
• utils: folder containing all helper functions
• privacy-gan.yml: conda environment to be able to run all scripts

How to run

1. Install the conda environment using the privacy-gan.yml file
2. Specify configurations and paths in config.py
3. To run the GAN with differential privacy, specify the hyperparameters, e.g.:

   python3 run_gan.py \
       --trial "GAN1" \               # trial name
       --noisem 1 \                   # noise multiplier value
       --max_norm_dp 1 \              # value for DP noise clipping
       --clip_param_W 0.01 \          # value for WGAN clipping
       --n_discr 5 \                  # number of D updates
       --batch_size 32 \              # batch size
       --lrd 0.00005 --lrg 0.00005 \  # learning rate D and G
       --ndf 96 --ngf 96 \            # number of filters D and G
       --kd 4 --kg 4 \                # kernel size D and G
       --strd 2 --strg 2 \            # strides D and G
       --opg 0 --padg 1 \             # (output) padding G
       --epochs 50 \                  # number of epochs
       --seed 12                      # number for random seed

4. For generating image-labels from the trained generator, run the generate_samples.py-file, e.g.:

   python3 generate_samples.py \
       --trial "GAN1"                 # trial name 
       --epoch 49                     # determine epoch for generation
       --thresh 0.8                   # threshold value for binarizing the label mask
       --lrg 0.00005                  # learning rate G
       --strg 2                       # strides G
       --kg 4                         # kernel size G
       --padg 1                       # padding G
       --opg 0                        # output padding G

5. The U-net parameters are mostly specify in the config.py-file. Run the U-nets as follows:

   python3 run_unet.py \
       --trial "GAN1"                 # trial name
       --epoch 49                     # determine epoch for generation
       --thresh 0.8                   # threshold value for binarizing the label mask

6. Evaluate the U-nets like this:

   python3 eval_unet.py \
       --trial "GAN1"                 # trial name
       --epoch 49                     # determine epoch for generation
       --thresh 0.8                   # threshold value for binarizing the label mask

How to cite

Kossen, T., Hirzel, M.A., Madai, V.I., Boenisch, F., Hennemuth, A., Hildebrand, K., Pokutta, S., Sharma, K., Hilbert, A., Sobesky, J., Galinovic, I., Khalil, A.A., Fiebach, J.B., Frey, D., 2022. Toward Sharing Brain Images: Differentially Private TOF-MRA Images With Segmentation Labels Using Generative Adversarial Networks. Frontiers in Artificial Intelligence 5. doi: 10.3389/frai.2022.813842.