基于深度学习的头颈部放疗计划图像分割新方法的研究

头颈部CT图像肿瘤靶区和放疗危及器官的准确分割，是合理制订头颈部放疗计划、提高放疗质量、减少并发症的关键。因现有理论和方法的局限及头颈部肿瘤和危及器官的特殊性，对其全自动、高精度的分割问题长期未能得到成功解决。本项目将针对这一问题开展深入的研究，提出新理论、新方法，弥补前人工作的不足，实现对CT图像肿瘤靶区和危及器官的全自动、高精度分割，分割精度和速度将达到临床实用的程度。项目研究将建立包含300个病例的标准数据集和包含1000个病例的弱标注数据集，用于头颈部分割网络的训练与测试。项目拟用自动机器学习方法，为分割目标分别建立最优网络结构、数据增强策略及数据预处理方法；拟通过点云预处理网络学习分割目标的个性化三维结构信息，大幅提升分割网络的整体性能；将借助弱标注数据并结合迁移学习方法，提升分割网络的分割精度与泛化能力。研究成果可提高头颈部放疗的质量，减轻医生负担，同时具有重要的学术价值。

For radiation therapy of head-and-neck tumor, accurate segmentation of the gross target volume of the tumor and the Organ At Risk (OAR) surrounding the tumor in CT images is the key to make a proper treatment plan and improve the quality and reduce the complications of the therapy. However, because of the limitation of existing theory and methods in this field and the characteristics of the head-and-neck tumor and related OARs, fully automatic segmentation with high accuracy has not been achieved. In this project, we will conduct in-depth research and develop new theory and methods to solve the problem existing in current studies, so that we can achieve fully automatic segmentation of the tumor and the OARs with a level of accuracy that is high enough for clinical use. We will establish a data set consisting of 300 cases with accurate labeling and 1000 cases with weak labeling of head-and-neck tumor and OARs, and we will use these data for training deep neural networks. We will utilize automatic machine learning techniques to find the best network architecture, the best data augmentation strategy and the best preprocessing method for each target to be segmented. We will use point cloud processing network to learn organ-specific 3D structure of each target and utilize them to increase the segmentation accuracy. In addition, we will make use of the weak labeling data by using transfer learning to further improve the accuracy and generalization capability of the segmentation network. The results of this project will help improve the quality of head-and-neck radiation therapy and reduce of the workload of radiologists, and they also have important academic values.