Computer Vision News - January 2021

18 Finally, for a slightly different approach nnU-Net also offers the possibility to merge the predictions calculated by two different architectures. The code below will run the inference using the 2D UNet and the 3D full resolution one and merge the predictions saving them into the folder called OUTPUT_FOLDER_ENSEMBLE. The Tool of the Month Training with nnU-Net Training a new model with nnU-Net is similarly easy! It is based on construction of a dataset fingerprint (a set of dataset-specific properties such as image sizes, voxel spacings, intensity information etc) which is extrapolated from the raw images given. This is used to build an ad-hoc segmentation pipeline configuration for the dataset. The code below performs the pre-processing for task 005 which corresponds to the Pancreas task. For other tasks, this code will be different and dependent on the folder placed inside the raw data directory. nnUNet_predict -i $nnUNet_raw_data_base/nnUNet_raw_data/Task005_Prostate/imagesTs/ -o OUTPUT_DIRECTORY_3D -t 5 --save_npz -m 3d_fullres nnUNet_predict -i $nnUNet_raw_data_base/nnUNet_raw_data/Task005_Prostate/imagesTs/ -o OUTPUT_DIRECTORY_2D -t 5 --save_npz -m 2d nnUNet_ensemble -f OUTPUT_DIRECTORY_3D OUTPUT_DIRECTORY_2D -o OUTPUT_FOLDER_ENSEMBLE -pp POSTPROCESSING_FILE nnUNet_plan_and_preprocess -t 005 --verify_dataset_integrity