For more details, check out:

• training/data_augmentation/default_data_augmentation.py
• training/network_training/nnUNetTrainerV2.py
• training/data_augmentation/data_augmentation_moreDA.py

Default 3D augmentation params

• Rotation x, y, z: (-30 degree, +30 degree)

• Scaling: (0.7, 1.4)

• Gamma: (0.7, 1.5)

• Mirror

• Gaussian noise

• Gaussian blur

• Brightness

• Contrast

• Low resolution simulation

• Dummy 2D aug: depends on the plans from plan_and_preprocess command

• No elastic deformation

• No random crop

Default 2D augmentation params

Almost the same as 3D augmentation with some minor changes.

• No dummy 2D augmentation
• No rotation of y, z axes

Dummy 2D augmentation

Used only in 3D augmentation, 3D data are converted to 2D, applied some 2D augmentations then converted back to 3D.