SpMVNet: Spatial Multi-view Network for Head and Neck Organs at Risk Segmentation

Hongzhi Liu, Shiyu Zhu, Qianjin Feng, Yang Chen

Abstract: Head and neck (HaN) cancers are often treated with radiotherapy. Since radiation inevitably causes damage to human organs, it is necessary to control the dose of radiation in different areas during radiation therapy to protect organs at risk (OARs). To solve these incompatible problems, we proposed an end-to-end spatial multi-view network for head and neck organs at risk segmentation, named SpMVNet, to take advantage of both spatial continuous context and multi-view relevance in whole volume CT images. The proposed method includes a symmetric segmentation network (SymNet) and a continuous context network (CCNet), making full use of organs’ structural symmetry in CT slices and spatial contextual information of volume data. Our proposed method is validated on the MICCAI 2015 Head and Neck Automatic Segmentation Challenge datasets. Extensive experiments show that it achieves lower error range for most organ segmentation with better evaluation metrics than state-of-the-art methods. This proposed method is helpful to improve the precision of organ segmentation in radiotherapy.