面向肺病诊断的CT图像三维交织管树结构分割与重建

The tubular tree structures like vasculture and bronchi, which are commonly seen in human body, have received wide attention in medical imaging due to their utilization in anatomical localization and pathological diagnosis. To satisfy the requirement of Computer-Aided Diagnosis (CAD) of pulmonary diseases, this project focuses on the segmentation and reconstruction of pulmonary airway, artery and vein trees, where a series of algorithms are presented for low-level feature enhancement, skeleton extraction and tube wall segmentation. Firstly, to solve the problem of adhering tubes separation, a new medialness likelihood function will be constructed with intercept-lines sampling from the fabric tensor, based on an observation that the boundary gradient vectors usually orient toward the tube centers. Then, using a regional geodesic voting scheme and some 3D shape descriptors, an axial connectivity metric could be defined by utilizing the global information to compensate the abnormalities like local data missing, bifurcation and deformations. Simultaneously, by investigating a scheme to merge different prior knowledge such as anatomy, shape and geometry, both a streamline vector field and a cost function would be generated to divide and reconstruct different kinds of tree skeletons. Finally, under a geometric deformable model framework, a new front-propagation speed could be formulated by integrating the homology between adjacent branches and the fractal property along the trees to realize an accurate location and segmentation of the tube walls. Our research would be beneficial to the extraction of pathological indices and localization of lesions in CT images, and is expected to provide a core technology support for development of lung CAD systems.

以血管和支气管为代表的树状结构是人体内常见组织，因其解剖定位和病理诊断价值在医学图像领域受到广泛关注。本项目结合肺部疾病计算机辅助诊断（CAD）需求，对CT图像中三维肺气道树、肺动脉和肺静脉树的分割与重建问题展开研究，分别从底层特征增强、骨架提取、管壁分割等方面提出一系列算法。首先，基于边界梯度向量向心性观察，融合构件张量截取线抽样，构建新的中轴似然度量以解决粘连管形分离难题；然后，借助区域测地线投票和3D空间形状描述算子，有效融合全局信息形成兼容局部数据缺失或分叉形变的轴向连接度测量；同时，探索解剖、形状和几何等多种先验知识融合机制，构造流线向量场和代价函数，对不同管树骨架进行分离和重建；最后，利用邻接管树对偶关系和分形相似性定义波阵面演化速度，在几何形变模型框架下实现管壁定位和分割。本项目研究成果有助于CT影像病理参数提取和病灶定位，可望为肺病CAD系统开发提供核心技术支持。

以血管和支气管为代表的树状结构是人体内常见组织，因其解剖定位和病理诊断价值在医学图像领域受到广泛关注。 本项目结合肺部疾病计算机辅助诊断(CAD)需求，主要研究肺部CT图像中三维管树解剖结构检测和分割方法：（1）针对粘连血管检测难题，提出一种基于改进的梯度向量流的管状中心增强滤波器。（2）为了分割细小血管，提出一种各向异性连续最大流分割方法，即利用血管轴向方向构建各向异性正则项。（3）在原有窄条微分滤波器的基础上融入肺裂方向信息，并提出方向区域剖分和迭代融合的肺裂分割方法。（4）借助气管、肺动脉解剖结构信息定位肺裂区域，联合肺裂增强、曲面拟合方法，实现肺裂自动分割。（5）将方向窄条微分滤波器扩展成方向平板微分滤波器，可直接从三维空间中增强和分割微弱肺裂，具有分割低剂量肺部CT图像肺裂的能力。（6）借助肺裂、气管和血管检测的研究成果，在分水岭割框架下实现了肺叶自动分割。