肺癌是目前世界范围内致死率最高的癌症之一，对肺癌的早发现、早治疗能够显著提高患者的存活率。医学影像检查是肺癌筛查诊断的重要依据，利用人工智能方法实现病变的自动检测分析能够辅助放射科医生准确快速诊断病情，避免繁重的阅片工作，减少漏检与误检。本论文在国家重点研发计划、国家自然科学基金等项目的支持下，以肺癌病变肺结节检测为目标，开展CT影像肺组织结节语义特征识别方法研究，包括：区域边缘协作式水平集图像分割框架、协作式多相水平集模型构建方法、协作式几何活动轮廓肺组织CT自动分割、以及面向CT影像语义特征的肺结节识别方法。论文主要内容及研究成果如下：（1）针对混合水平集模型在区域与边缘信息协作方面缺乏理论依据、能量权重系数难以选择的问题，提出了区域边缘协作式水平集图像分割框架。它能够利用现有水平集模型的近似等价形式构建区域边缘混合模型，其能量权重可根据模型的全局最优条件进行选择。同时，该框架能够使区域边缘信息内在地相互协作，提高图像分割的精度、计算效率以及对水平集函数初始化的鲁棒性。（2）针对多相水平集模型等价近似误差较大、每个能量项对整体能量泛函的贡献很难定量描述、参数更加难以确定等问题，提出了协作式多相水平集模型构建方法。该方法利用现有模型的完全等价形式构建协作式多相水平集模型，根据其全局最优条件可方便地选择模型参数。该方法定量描述了区域信息与边缘信息的内在关系，使图像分割的精度、计算效率及稳定性得以提高。（3）针对经典水平集模型对肺组织模糊边界及狭长区域分割精度较低、对连续CT序列分割稳定性较差等问题，提出了一种全局优化协作式几何活动轮廓模型。该模型将边缘信息与区域信息相融合，提高了肺组织的分割精度。同时，模型的全局最优性使其能够充分利用相邻图像之间的先验解剖结构信息，显著提高连续胸腔CT序列肺组织分割的稳定性与计算效率。（4）针对肺结节语义特征提取假设条件存在的局限性，提出了一种像素移除率语义特征对肺结节进行量化表示，有效区分肺结节与肺血管。同时，在像素移除率特征的基础上构建了一套面向CT影像语义特征的肺结节检测系统，能够有效识别孤立结节、血管粘连结节与胸腔壁结节，为面向CT影像的肺癌筛查与检测提供可靠技术支撑。

Lung cancer is currently one of the most lethal diseases in the world. Early detection and treatment of lung cancer can greatly improve the survival rate of patients. The examination by medical imaging is an important basis for diagnosis of lung cancer. Automatic detection and analysis of lesions by artificial intelligence can assist radiologists in accurately and quickly diagnosing the disease, avoiding heavy workload to check numerous images, and reducing the number of false and missed detection. With the support of the National Key Research and Development Program and the National Natural Science Foundation of China, this dissertation focuses on pulmonary nodules detection caused by lung cancer and contributes in semantic features enabled recognition methodology for lung nodules on thoracic CT Images, which includes the following points: region and edge synergetic level set framework for image segmentation, a generic approach for constructing synergetic multiphase level set models, automatic lung segmentation on CT images using a synergetic geometric active contour model, and the detection of lung nodules based on semantic features of CT images. The main content and research achievements of this dissertation are as follows:Firstly, in hybrid level set models, the theoretical basis of the collaboration mechanism between the region and edge information is insufficient, which makes it difficult to select appropriate energy weights. Thus, a region and edge synergetic level set framework is proposed for image segmentation, which makes it possible to use the approximate equivalent form of some existing level set models to construct hybrid ones containing region and edge information. Their energy weights can be selected based on the global optimization condition deduced from the framework. Meanwhile, the framework can intrinsically make the region and edge information cooperate with each other, which improves the segmentation accuracy, computational efficiency, as well as the robustness to the initialization of level set functions.Secondly, for the multiphase level set model, the equivalent approximation error is large, the contribution of each energy term to the overall energy functional is difficult to describe quantitatively, and the parameters are more difficult to determine. Therefore, a generic approach for constructing synergetic multiphase level set models is proposed. This method uses the fully equivalent form of the existing models to construct synergetic multiphase level set models, which makes it easy to select energy weights based on the global optimization condition of the proposed approach. This method quantitatively describes the inherent relationship between region and edge information, which improves the accuracy, computational efficiency and stability of image segmentation.Thirdly, lung segmentation using classical level set models causes low accuracy in weak boundaries and narrow bands, and poor stability when segmenting serial CT sequences. Thus, a global optimal synergetic geometric active contour model is proposed to tackle these issues. The model fuses both region and edge information, which improves the accuracy of lung segmentation. Meanwhile, the global optimality of the model allows it to make full use of the prior anatomical information between adjacent CT images, which significantly improves the stability and computational efficiency of lung segmentation in serial thoracic CT sequences.Finally, since there are some limitations of the hypothesis which is used to extract semantic features of lung nodules, a semantic feature named Pixels Eliminated Rate is presented to characterize lung nodules, which can effectively distinguish pulmonary nodules from lung vessels. Simultaneously, based on the proposed semantic feature, a detection system for lung nodules recognition on CT images is constructed, which can effectively identify solitary nodules, juxta-vascular nodules, and juxta-pleural nodules. The proposed system provides a reliable technical support for screening of lung cancer on CT images.