A Keras CycleGAN for nifti data

We provide a Keras implementation for unpaired and image-to-image translation, i.e. CycleGAN [1], for nifti data.

We proposed an application of CycleGAN to generate synthetic diffusion MRI scalar maps from structural T1-weighted images, see our paper [2].

Getting started

Prepare training data

We prepare the training data by stacking subjects on the fourth dimention.
For example, we extract 1 slice from each subject for 1000 subjects and then stack all slices to the fourth dimention.
The created training data will have the size of [X, Y, 1, 1000].

Training

# Create a CycleGAN on GPU 0 
myCycleGAN = CycleGAN(0) 

# Set directories  
trainA_dir = '/home/xuagu37/CycleGAN/data/T1_training.nii.gz'  
trainB_dir = '/home/xuagu37/CycleGAN/data/FA_training.nii.gz'  
models_dir = '/home/xuagu37/CycleGAN/train_T1_FA/models'  
output_sample_dir = '/home/xuagu37/CycleGAN/train_T1_FA/output_sample.png'  

# Set training parameters  
batch_size = 10  
epochs = 200  
normalization_factor_A = 1000  
normalization_factor_B = 1  

# Start training  
myCycleGAN.train(trainA_dir, normalization_factor_A, trainB_dir, normalization_factor_B, models_dir, batch_size, epochs, output_sample_dir=output_sample_dir, output_sample_channels=1)

Synthesize

# Set directory to the trained model  
G_X2Y_dir = '/home/xuagu37/CycleGAN/train_T1_FA/models/G_A2B_weights_epoch_100.hdf5'  

# Set directory to the test data  
test_X_dir = '/home/xuagu37/CycleGAN/data/T1_test.nii.gz'  

# Set directory to save the synthetic data  
synthetic_Y_dir ='/home/xuagu37/CycleGAN/train_T1_FA/synthetic/FA_synthetic.nii.gz'  

# Synthesize
normalization_factor_X = 1000   
normalization_factor_Y = 1  
myCycleGAN.synthesize(G_X2Y_dir, test_X_dir, normalization_factor_X, synthetic_Y_dir, normalization_factor_Y)

References