Computer Vision News - January 2021

16 The next step consists into setting up three folders, two containing the data and one for the results (model weights). • The 1 st path is called nnUNet_raw_data_base: this is where the subfolder containing the raw data should be added, which in turn contains another subfolder named after the task to perform. • The 2 nd path is the nnUNet_preprocessed and the pre-processed data will be saved here. The data will also be read from this folder during training. • The 3 rd path is the RESULTS_folder where the model weights from a new training are saved, or from pretrained models are downloaded. Now we should be in the main development nnU-Net folder and we can type the lines below in the terminal to create the necessary paths and set them up as reference folders: The last three line can be just run in the terminal for a single time use or saved in the bash profile to make them permanent. Inference with nnU-Net The Tool of the Month mkdir data cd data mkdir nnUNet_raw nnUNet_preprocessed nnUNet_trained_models export nnUNet_raw_data_base="./nnUNet_raw" export nnUNet_preprocessed="./nnUNet_preprocessed" export RESULTS_FOLDER= “./ nnUNet_trained_models" Now, what you need to do is just to move your data into the right location, which should be under nnUnet->data->nnUNet_raw->nnUnet_raw_data , and have a look at what the right format is. The general folder of the raw data should look like Fig. 4, while all the images in the training/testing/label folder must be 3D nifti files ( .nii.gz ) and have unique identifiers. Imaging files must follow the naming convention illustrated in Fig. 5 with case_identifier_XXXX .nii.gz (where XXXX is the modality identifier). For labels, the last is not necessary.