Computer Vision News - January 2021

The nnU-net was mentioned once more in our magazine, during an interview with Klaus-Maier-Hein . Both Lena and Klaus explained the meaning behind the name: nnU-Net stands for no-new-U-net , and it was created by Fabian Isensee and his co- authors (Klaus was last author). They followed an interesting principle: to not create anything new but optimise what we already have. This of course refers to the U-net architecture, which has been vastly used in the field and highly valued since its first presentation at MICCAI. nnU-Net is a deep learning-based segmentationmethod that automatically configures itself, including pre-processing, network architecture, training and post-processing for any new task in the biomedical domain, but it is only based on three relatively simple and state-of-the-art architectures: the 2D U-Net (Fig. 1), the 3D U-Net (Fig. 2) and the U-Net Cascade (Fig. 3). While the first two are probably already perfectly known by our readers, the last is a method which generates low resolution segmentations and subsequently refines them. It employs two 3DU-Nets, that first (Stage 1) downsample the data and upsample the resulting segmentation maps, and then concatenate the segmentations to refine them through another U-Net and obtain the full resolution data (Stage 2). So the obvious question here could be: then what’s the novelty about it? Well… nnU- net is a self-adapting framework. And what does this mean? The Tool of the Month 14 nnU-Net Dear readers, welcome back and (can we finally say it?) Happy New Year! We hope this year of interviews and reviews will be even more interesting than the previous one. And with that in mind, let’s start talking about the tool of the month: nnU-net, which was first mentioned in our June 2019 issue by Lena Maier-Hein as the winning algorithm of the Medical Segmentation Decathlon. This is a tool developed by the group in the Division of Medical Image Computing in the German Cancer Research Center inHeidelberg and of course inspired by and based on the 2015 well-known U-Net architecture. by Marica Muffoletto