Computer Vision News - January 2021

17 If we want to check that the pre-processed data has now been created, we can check that nnUnet->data-> nnUNet_preprocessed/Task005_Prostate is not empty and also note that the pre-processed data folder only contains the training cases, while for the test images, the pre-processing happens on the fly during inference. Now, we can run the training using the following command: But the above is just using one of the many options offered by nnU-Net. More generally, the arguments in the training command stand for: The types of CONFIGURATION available are: 2d, 3d_fullres, 3d_lowres, 3d_cascade_ fullres, while the FOLD argument refers to the cross validation ranging from 0 to 4- since nnU-Net uses a 5-fold cross validation. The weights of this model are saved in the RESULTS_FOLDER set up at the start, and they can be in turn used for inference as seen previously. Conclusions nnU-Net is a fast and out-of-the-box segmentation tool based on deep learning which was able to outperform several methods on different tasks, showing an incredible flexibility. This can be even further augmented by a user who wants to integrate it with independent pieces of code. During the interview with Lena Maier-Hein , computer science professor in the same university where nnU-Net was built, she declared than nnU-Net “is not about inventing something entirely novel, but it’s about doing things right. I like this hypothesis. I was really surprised and excited to see that the algorithm won by a large margin.” We find equally fascinating the idea behind this project and hope that the future can bring not only innovative things but also innovative ways and perspectives to make better use of what we already have. nnU-Net nnUNet_train 2d nnUNetTrainerV2 Task005_Prostate 1 --npz nnUNet_train CONFIGURATION TRAINER_CLASS_NAME TASK_NAME_OR_ID FOLD --npz (additional options) 9