nnU-Net: Self-adapting Framework for U-Net-Based Medical Image\n Segmentation

Abstract

The U-Net was presented in 2015. With its straight-forward and successful\narchitecture it quickly evolved to a commonly used benchmark in medical image\nsegmentation. The adaptation of the U-Net to novel problems, however, comprises\nseveral degrees of freedom regarding the exact architecture, preprocessing,\ntraining and inference. These choices are not independent of each other and\nsubstantially impact the overall performance. The present paper introduces the\nnnU-Net ('no-new-Net'), which refers to a robust and self-adapting framework on\nthe basis of 2D and 3D vanilla U-Nets. We argue the strong case for taking away\nsuperfluous bells and whistles of many proposed network designs and instead\nfocus on the remaining aspects that make out the performance and\ngeneralizability of a method. We evaluate the nnU-Net in the context of the\nMedical Segmentation Decathlon challenge, which measures segmentation\nperformance in ten disciplines comprising distinct entities, image modalities,\nimage geometries and dataset sizes, with no manual adjustments between datasets\nallowed. At the time of manuscript submission, nnU-Net achieves the highest\nmean dice scores across all classes and seven phase 1 tasks (except class 1 in\nBrainTumour) in the online leaderboard of the challenge.\n