cfgs/generator/train/shape_id.yaml

device_generator: cuda:1

split: train # train or test

dataset_names: ['ADHD'] # list of datasets
dataset_probs:
modality_probs: {
  'ADHD': {'T1': 0.5, 'T2': 0., 'FLAIR': 0., 'CT': 0., 'synth': 1.}, # healthy
  'HCP': {'T1': 0., 'T2': 0., 'FLAIR': 0., 'CT': 0., 'synth': 1.}, # healthy
  'AIBL': {'T1': 0.25, 'T2': 0.5, 'FLAIR': 0.75, 'CT': 0., 'synth': 1.}, # healthy
  'OASIS': {'T1': 0.3333, 'T2': 0., 'FLAIR': 0., 'CT': 0.6667, 'synth': 1.}, # healthy
  'ADNI': {'T1': 0.5, 'T2': 0., 'FLAIR': 0., 'CT': 0., 'synth': 1.}, # healthy / wmh
  'ADNI3': {'T1': 0., 'T2': 0., 'FLAIR': 0., 'CT': 0., 'synth': 1.}, # wmh
  'ATLAS': {'T1': 0.5, 'T2': 0., 'FLAIR': 0., 'CT': 0., 'synth': 1.}, # stroke
  'ISLES': {'T1': 0., 'T2': 0., 'FLAIR': 0.5, 'CT': 0., 'synth': 1.}, # isles
} 
mix_synth_prob: 0. # blend synth with real images
dataset_option: brain_id

# setups for training/testing tasks 
task:
  T1: False
  T2: False
  FLAIR: False
  CT: False
  pathology: True

  super_resolution: False
  segmentation: False 
  registration: False
  surface: False
  distance: False 
  bias_field: False
  contrastive: False


# setups for augmentation functions to apply
augmentation_steps: ['gamma', 'bias_field', 'resample', 'noise'] 

# setups for generator
generator:

  #size: [128, 128, 128]
  size: [100, 100, 100]

  photo_prob: 0.2
  max_rotation: 15
  max_shear: 0.2
  max_scaling: 0.2
  nonlin_scale_min: 0.03
  nonlin_scale_max: 0.06
  nonlin_std_max: 4
  bag_prob: 0.5
  bag_scale_min: 0.02
  bag_scale_max: 0.08
  bf_scale_min: 0.02
  bf_scale_max: 0.04
  bf_std_min: 0.1
  bf_std_max: 0.6
  gamma_std: 0.1
  noise_std_min: 5
  noise_std_max: 15
  exvixo_prob: 0.25
  exvixo_prob_vs_photo: 0.66666666666666

  pv: True
  random_shift: False
  deform_one_hots: False
  integrate_deformation_fields: False
  produce_surfaces: False
  bspline_zooming: False  
  n_steps_svf_integration: 8
  nonlinear_transform: True

  ct_prob: 0
  flip_prob: 0. 
  
  pathology_prob: 1. # pathology_prob when synth
  random_shape_prob: 1. # initialize pathol shape from random noise (v.s. existing shapes)
  augment_pathology: True


  # brain-id customized setups

  # mild-to-severe intra-subject aug params
  mild_samples: 2
  all_samples: 4
  all_contrasts: 4 # >= 1, <= all_samples
  num_deformations: 1




pathology_shape_generator:
  perlin_res: [2, 2, 2] # shape must be a multiple of res
  mask_percentile_min: 85
  mask_percentile_max: 99.6
  integ_method: dopri5 # choices=['dopri5', 'adams', 'rk4', 'euler'] 
  bc: neumann # choices=['neumann', 'cauchy', 'dirichlet', 'source_neumann', 'dirichlet_neumann'] 
  V_multiplier: 500
  dt: 0.1
  max_nt: 10 # >= 2
  pathol_thres: 0.2
  pathol_tol: 0.000001 # if pathol mean < tol, skip




# brain-id customized setups

mild_generator:   
  bag_prob: 0.1
  bag_scale_min: 0.01
  bag_scale_max: 0.02
  bf_scale_min: 0.01
  bf_scale_max: 0.02
  bf_std_min: 0.
  bf_std_max: 0.02
  gamma_std: 0.01
  noise_std_min: 0.
  noise_std_max: 0.02 


severe_generator:   
  bag_prob: 0.5
  bag_scale_min: 0.02
  bag_scale_max: 0.08
  bf_scale_min: 0.02
  bf_scale_max: 0.04
  bf_std_min: 0.1
  bf_std_max: 0.6
  gamma_std: 0.1
  noise_std_min: 5
  noise_std_max: 15