Exploring advanced architectural variations of nnUNet

Abstract

The nnUNet is a state-of-the-art deep learning based segmentation framework which automatically and systematically configures the entire network training pipeline. We extend the network architecture component of the nnUNet framework by newly integrating mechanisms from advanced U-Net variations including residual, dense, and inception blocks as well as three forms of the attention mechanism. We propose the following extensions to nnUNet, namely Residual-nnUNet, Dense-nnUNet, Inception-nnUNet, Spatial-Single-Attention-nnUNet, Spatial- Multi-Attention-nnUNet, and Channel-Spatial-Attention-nnUNet. Furthermore, within Channel-Spatial- Attention-nnUNet we integrate our newly proposed variation of the channel-attention mechanism. We demonstrate that use of the nnUNet allows for consistent and transparent comparison of U-Net architectural modifications, while maintaining network architecture as the sole independent variable across experiments with respect to a dataset. The proposed variants are evaluated on eight medical imaging datasets consisting of 20 anatomical regions which is the largest collection of datasets on which attention U-Net variants have been compared in a single work. Our results suggest that attention variants are effective at improving performance when applied to tumour segmentation tasks consisting of two or more target anatomical regions, and that segmentation performance is influenced by use of the deep supervision architectural feature.