Automatic segmentation of gross target volume of nasopharynx cancer using ensemble of multiscale deep neural networks with spatial attention

Abstract Radiotherapy is the main treatment method for nasopharynx cancer. Delineation of Gross Target Volume (GTV) from medical images a prerequisite radiotherapy. As manual delineation time-consuming and laborious, automatic segmentation GTV has potential to improve efficiency this process. This work aims automatically segment cancer Computed Tomography (CT) images. However, it challenged by small target region, anisotropic resolution clinical CT images, low contrast between region surrounding soft tissues. To deal with these problems, we propose 2.5D Convolutional Neural Network (CNN) handle different in-plane through-plane resolutions. We also spatial attention module enable network focus on target, use channel further performance. Moreover, multi-scale sampling training so that networks can learn features at scales, which are combined multi-model ensemble robustness results. estimate uncertainty results based our model ensemble, great importance indicating reliability radiotherapy planning. Experiments 2019 MICCAI StructSeg dataset showed (1) Our proposed better performance than commonly used 3D networks. (2) mechanism make pay more accuracy. (3) The achieves robust results, simultaneously obtain information indicate mis-segmentations decisions.