血管的识别是临床和生理学上的诊断治疗和受试者间比较的先决任务。然而，血管系统的异构性增加了外科医生的负担。利用计算机辅助系统解决受试者的标记解剖血管名称问题对于血管及器官疾病的诊断、治疗和病例分析是至关重要的。 血管结构的分支模式的拓扑异质性很大，例如：血管的顺序、结构缺失或多余的假阳性结构等，理解并标记血管是困难的。因此，手动标记这些血管结构是耗时且容易出错的。除了这些生理学的困难外，在自动标记血管的任务中，由于图像预处理的几个必要步骤不可避免地会引入额外的困难而更具挑战性，例如造影伪像噪声、分割误差和不准确的描述符等。 本文提出一种三维血管标记任务的解决框架。将受试者的骨架线模型作为输入，这意味着它可以适用于不同成像数据中血管数据的解剖学的和不可预测的异质性。框架首先从血管网络的骨架结构中提取分支，基于拟合的参数化曲线获取与全局坐标和尺度无关的一组几何特征，而后利用多分类器预测目标血管树中的分支在候选标签集合上的似然分布。最后，针对不同的血管网络元素和结构设计了不同的拓扑约束方法，并基于具有约束变换策略的隐马尔可夫模型获取全局拓扑约束下的最优标记序列。 本文的贡献主要有如下几个方面： 1）通用的血管标记任务框架：本文为研究者提供三维血管自动标记任务的框架。主要流程为首先基于局部几何特征设计多分类器得到标记对象的概率分布，然后基于血管网络的结构建立拓扑一致性约束，最后使用拓扑约束的概率图模型以提高血管标记的准确性。 2）基于局部几何特征的自动标记：本文利用从原始影像分割提取得到的三维骨架线作为输入，利用B样条拟合离散骨架得到连续参数化曲线，基于此提取与坐标无关的局部几何特征，使得数据不需要预先进行严格对齐，进而使用集成学习方法构建基于局部几何特征的似然概率预测模型，具有比传统方法更加准确的分类效果。 3）基于全局拓扑约束的自动标记：将拓扑识别转化为由Viterbi算法求解的最优空间序列问题。本文针对不同的目标结构设计不同的序列构建方法，能够减少拓扑异质性，更好的反映血管结构的层次和邻接关系，并且能够处理未修剪的数据。 4）实验结果：通过在50套脑动脉环的公开数据和37套腹部动脉临床数据的交叉验证来评价本文方法，脑动脉的标记正确率达到了90.46%，腹部动脉的标记正确率达到91.94％。 本文提出了一种新颖的血管自动标记方法框架，并将方法应用在脑动脉环和腹部动脉结构，以此验证本研究方法的可行性和适用性，具有重要的理论意义和实际应用价值。

Identification of blood vessels is a prerequisite for clinical and physiological diagnostic treatment and comparison among inter-patients. However, the heterogeneity of the vascular system increases the burden on the surgeon. Using computer-aided systems to address the anatomical vessel labeling problem is critical for the diagnosis, treatment, and medical case analysis of vascular and organ diseases. Due to the large topological heterogeneity of branching pattern of human vascular structures, such as the order of branches, the absence of branches or the presence of redundant false positive structures, etc., it is difficult to understand and label the vascular structure. Therefore, manual labeling of these vascular structures is time-consuming and error-prone. Set aside these physiological difficulties, the automatic labeling task is even more challenging, because several necessary steps for a labeling process introduce additional difficulties, such as imaging artifacts and noises, segmentation errors, and inaccurate characterization. This paper presents a framework for 3D vascular labeling tasks. The subject’s skeleton model is taken as input, this means it can tolerate the anatomical and unpredictable heterogeneity of vascular data in different imaging types. Branches are extracted from the skeleton structure of the vascular network. A set of geometric features of branches are acquired which are independent of global coordinates and scales based on the fitted parametric curve. Then the multi-classifier is used to predict the likelihood distribution in candidate label set of all branches in the target vessel tree. Finally, different topological constraint methods are designed for different vascular network elements and structures. Based on the hidden Markov model with the constrained transformation strategy, the optimal label sequence that satisfies global topology constraints is obtained. The contributions of this paper are mainly as follows: 1) Universal vascular labeling framework: This article provides researchers with a framework for solving automatic labeling of 3D vascular models. The pipeline is to design a multi-classifier based on local geometric features to obtain the probability distribution of objects firstly, then establish the topological consistency constraint based on the vascular network structure, finally use probability graphical model with the topologically constraints to improve the accuracy of the vascular labeling. 2) Automatic labeling based on local geometric features: This paper uses the three-dimensional skeleton extracted from the original image segmentation as input, and uses B-splines to fit the discrete skeleton to obtain continuous parameterized curves. Based on these curves, coordinate-independent local geometric features are extracted, which alleviate the need for strict global alignment in advance, and then the ensemble learning method is applied as a likelihood probability prediction model based on these features, which has a more accurate classification accurate than the traditional methods. 3) Automatic labeling based on global topology constraints: Topological recognition is transformed into the problem of finding optimal spatial sequences, which could be solved by the Viterbi algorithm. Different sequence reconstruction methods are designed for different target structures to reduce topological heterogeneity, which better reflect the vascular level structure and adjacency relationship, and tolerate unpruned data. 4) Experimental results: This approach has been evaluated through the cross-validation method on 50 subjects of cerebral arteries public data and 37 subjects of abdominal arteries clinical data, and the accuracy of the labeling of cerebral arteries is 90.46%, and the accuracy in abdominal arteries is 91.94%. In this paper, a novel framework for automatic labeling of blood vessels is proposed, and the method is applied to the structure of cerebral arteries and abdominal arteries to verify the feasibility and applicability. It has important theoretical contributions and practical application significance.