Preprocessing

• Preprocess images with GenericPreprocessor.preprocess_test_case for 3D and PreprocessorFor2D.preprocess_test_case for 2D
• Crop non-zero regions
• Transpose based on predefined plan generated by nnUNet_plan_and_preprocess
• Resample:
• Normalize:

Predicting

• Run prediction of 5 folds and then calculate the mean of softmax values
• Save segmentation nifti from this mean softmax values
• For each fold:
  • Prediction on patches with Test Time Augmentation - mirroring patches on all axes then get the mean values
  • If more than 1 patches, use Gaussian importance map to merge patch prediction

Postprocessing

See postprocessing page