No New-Net

In this paper we demonstrate the effectiveness of a well trained U-Net in the context of the BraTS 2018 challenge. This endeavour is particularly interesting given that researchers are currently besting each other with architectural modifications that are intended to improve the segmentation performance. We instead focus on the training process, argue that a well trained U-Net is hard to beat and intend to back up this assumption with a strong participation in this years BraTS challenge. Our baseline U-Net, which has only minor modifications and is trained with a large patch size and a dice loss function already achieves competitive dice scores on the BraTS2018 validation data. By incorporating region based training, additional training data and a simple postprocessing technique, we obtain dice scores of 81.01, 90.83 and 85.44 and Hausdorff Distances (95th percentile) of 2.54, 4.97 and 7.