Integrating Residual, Dense, and Inception Blocks into the nnUNet

Abstract

The nnUNet is a fully automated and generalisable framework which automatically configures the full training pipeline for the segmentation task it is applied on, while taking into account dataset properties and hardware constraints. It utilises a basic UNet type architecture which is self-configuring in terms of topology. In this work, we propose to extend the nnUNet by integrating mechanisms from more advanced UNet variations such as the residual, dense, and inception blocks, resulting in three new nnUNet variations, namely the Residual nnUNet, Dense-nnUNet, and Inception-nnUNet. We have evaluated the segmentation performance on eight datasets consisting of 20 target anatomical structures. Our results demonstrate that altering network architecture may lead to performance gains, but the extent of gains and the optimally chosen nnUNet variation is dataset dependent.