We use our nnU-Net (“no-New-Net”) framework to fully automatically configure the entire segmentation pipeline used in this submission. All design choices such as preprocessing, architecture design, training, inference and postprocessing were determined fully automatically without user intervention. The final submission was generated by training an ensemble of a 2D and a 3D vanilla U-Net with minor architectural modifications.
The nnU-Net framework is not published yet as it is currently being evaluated in the context of the Medical Segmentation Decathlon challenge. Our paper will appear on arXiv after August 31st 2018.