Kang, E., Koo, H. J., Yang, D. H., Seo, J. B., & Ye, J. C. (2018). Cycle Consistent Adversarial Denoising Network for Multiphase Coronary CT Angiography. arXiv preprint arXiv:1806.09748.

• CYCLE_IDENTITY_GAN

• paper : https://arxiv.org/pdf/1806.09748.pdf
• reference code:
  • cyclegan : https://github.com/xhujoy/CycleGAN-tensorflow

• Input data Directory
  • DICOM file extension = ['.IMA', '.dcm']

$ os.path.join(dcm_path, patient_no, [LDCT_path|NDCT_path], '*.' + extension)

• Generator(revised image)

reference : Kang, E., Chang, W., Yoo, J., & Ye, J. C. (2018). Deep convolutional framelet denosing for low-dose ct via wavelet residual network. IEEE transactions on medical imaging, 37(6), 1358-1369.

1. To reduce network complexity, images are used directly as inputs to the network instead of the wavelet transform coefficients
   

2. The First convolution layer uses 128 set of 3 * 3 convolution kernels to produce 128 channel feature maps.

3. We have 6 set of module composed of 3 sets of convolution, batch normalization, and ReLU layers, and one bypass connection with a ReLU layer. Convolution layers in the modules use 128 set of 3 * 3 * 128 convolution kernels.

4. In addition, the proposed network has a concatenation layer that concatenates the inputs of each module and the output of the last module.

5. followed by the convolution layer with 128 set of 3 * 3 * 896 convolution kernels. The last convolution layer uses 15 sets of 3 * 3 * 128 convolution kernels. Finally, we add an end-to-end bypass connection.

• discriminator
  

• it consists of 5 convolution layers including the last fully-connected layer. The first convolution layer uses 64 sets of 4 * 4 convolution kernels, and the number of convolution kernels in the following layers is twice that of the previous layer except the last fully connected layer.

• mini batch size : 10
• patch size : 56*56
• opt : Adam(learning rate = 0.0002, beta1 = 0.5, beta2 = 0.999)
• learning rate decay : first 100 eppoch 0.0002 and linearly decreased it to zero over the next epochs.
• lambda(weight of cycle loss) : 10
• gamma(weight of identity loss) : 5
• epoch : 160
• normalized input : -1 ~ 1

• Directory

• dcm_path : dicom file directory
• LDCT_path : LDCT image folder name
• NDCT_path : NDCT image folder name
• test_patient_no : test patient id list(p_id1,p_id2...) (train patient id : (patient id list - test patient id list)
• checkpoint_dir : save directory - trained model
• test_npy_save_dir : save directory - test numpy file

• Image info

• patch_size : patch size
• whole_size : whole size
• img_channel : image channel
• img_vmax : max value
• img_vmin : min value

• Train/Test

• model : red_cnn, wgan_vgg, cyclegan, cycle_identity (for image preprocessing)
• phase : train | test

• Training detail

• end_epoch : end epoch (default = 160)
• lr : learning rate (default=0.0002)
• batch_size : batch size (default=10)
• L1_lambda : weight of cyclic loss (default=10)
• L1_gamma : weight of identity loss (default=5)
• beta1 : Adam optimizer parameter (default=0.5)
• beta2 : Adam optimizer parameter (default=0.999)
• ngf : # of generator filters in first conv layer
• nglf : # of generator filters in last conv layer
• ndf : # of discriminator filters in first conv layer

• others

• save_freq : save a model every step (iterations)
• print_freq : print frequency (iterations)
• continue_train : load the latest model: true, false
• gpu_no : visible devices(gpu no)
• unpair : unpaired image(cycle loss), (default=True)

• train

python main.py

• test

python main.py --phase=test