Model Configuration Readme

2022

Workflow

1. Data Preparation

   • Use preprocessing/generateSlicesForNetwork3D.py to package images as slices.
     • Creates numpy slices for model input during training.
     • Saves metadata containing dataset extremes (min and max values).
     • Saves image IDs.
   • Optionally, create a smaller batch of slices for test images to be generated every X epochs to check visual quality.

2. Training

   • Run train_3D.py to train the model.

Config

• input_seq: The input sequence for the model.

• output_seq: The output sequence for the model.

• database: Specifies the database to be used.

  • Options: "Heidelberg" or "PEGASUS"

• data: Specifies the shorthand for the database.

  • Default: ["hdb"] for Heidelberg, ["peg"] for PEGASUS

• model_path: Directory path where models are saved.

• gen_img_path: Directory path where generated images are saved.

• extremesDate: Date on which the file with min/max values of the cohort was created.

• gen_input_img_path_3D: Path for input images used during generation.

• gen_output_img_path_3D: Path for output images used during generation.

• extremesFile: Filename for the extremes file.

• specificImgsToGen: Specific images to generate during training for visual checks.

Model Settings

• save_model: Save model

• save_every_X_epoch: Frequency of saving the model.

• generate_while_train: Generate slices during model training.

• nr_imgs_gen: Number of images to generate (up to five).

• sliceToShow: Slice to generate as an example of model progression.

Training Settings

• epochs: Number of training epochs.

  • Default: 80

• paddingMode: Padding mode for convolutions.

  • Default: "reflect"

• model_name: Name of the model.

  • Default: 'pix2pix-pytorch'

• upsampling: Use upsampling instead of convtranspose.

  • Default: False

• batch_size: Size of the training batches.

  • Default: 8

• ngf: Number of generator filters.

  • Default: 256

• nr_layer_g: Number of layers in the generator (UnetG).

  • Default: 7

• nr_layer_d: Number of layers in the discriminator (PatchD).

  • Default: 3

• ndf: Number of discriminator filters.

  • Default: 64

• niter: Number of iterations.

  • Default: 10

• lr_g: Learning rate for the generator.

  • Default: 0.0001

• lr_d: Learning rate for the discriminator.

  • Default: 0.00001

• dropout_rate: Dropout rate.

  • Default: 0

• weightinit: Weight initialization method.

  • Default: "normal"

• init_sd: Initial standard deviation for normal distribution.

  • Default: 0.05

• init_mean: Initial mean for normal distribution.

  • Default: 0

• slope_relu: Slope for ReLU.

  • Default: 0.2

• patchD: Use patch discriminator.

  • Default: True

• kernel_size: Kernel size (only for no patchD).

  • Default: 5

• netG_type: Type of generator network.

  • Default: "unet"

• pretrainedG: Use pretrained generator (only for UnetG).

  • Default: False

• feature_matching: Use feature matching (only possible when no patchD).

  • Default: False

• label_smoothing: Use label smoothing.

  • Default: False

Loss Functions

• criterion_d: Loss criterion for the discriminator.

  • Default: "BCE"

• criterion_g: Loss criterion for the generator.

  • Default: "BCE"

• loss_ratio: Ratio for updating loss.

  • Default: False

• reconstruction_loss: Type of reconstruction loss.

  • Default: "L1-SSIM"

Optimizers

• optimizer_d: Optimizer for the discriminator.

  • Default: "adam"

• beta1_d: Beta1 parameter for the discriminator's optimizer.

  • Default: 0.5

• beta1_g: Beta1 parameter for the generator's optimizer.

  • Default: 0.5

• beta2_d: Beta2 parameter for the discriminator's optimizer.

  • Default: 0.999

• beta2_g: Beta2 parameter for the generator's optimizer.

  • Default: 0.999

GAN Architecture

• WGAN: Use Wasserstein GAN.

  • Default: False

• lbd: Lambda parameter for WGAN.

  • Default: 10

• nr_d_train: Number of discriminator training steps per generator step.

  • Default: 1

Continued Training

• continue_training: Continue training from a saved checkpoint.

  • Default: False

• load_from_epoch: Epoch number to load the model from.

  • Default: 20

Evaluation

• metrics: List of evaluation metrics.

  • Default: ["MSE", "NRMSE", "PSNR", "SSIM", "MAE"]

• save_nii: Save results in NIfTI format.

  • Default: True

Other

• generate_images_3D_inference.py: Run model in inference mode for test set.