【作者】 别继松；

【导师】 王珂；

【作者基本信息】 吉林大学 ， 信号与信息处理， 2012， 硕士

【摘要】 基于支持向量机的肺部结节计算机断层扫描(Computed Tomography，CT）图像的分割与识别是肺部计算机辅助诊疗(Computer-Aid Diagnosis，CAD）的核心，能够提供精简准确信息，提高医师诊断效率。准确的图像分割方法、合理的图像特征提取和泛化能力较强的机器学习方法是计算机辅助诊疗的关键。针对图像分割与识别问题的国内外研究重点如下：1.肺部结节提取方面，研究重点是肺结节的特性分析。根据图像的特点和肺部结节的特点使用针对性较强的方法分别提取孤立型肺部结节和粘连肺壁的肺部结节。2.肿瘤识别方面，研究重点是如何选择合适的图像特征，通过统计学理论，设计期望风险尽可能小的识别分类系统。针对以上问题，本文的主要研究工作如下：1.根据计算机断层扫描图像的成像特点和肺部图像的特点，使用基于改进滚球法的分割方法对粘连肺壁的肺部结节区域进行提取；使用区域增长法对肺区内孤立型结节区域进行提取。提取结果作为后续识别的学习样本库。实验结果表明，改进的滚球法可以有效提取粘连肺壁的结节，基于区域增长的方法可以有效提取独立型结节。2.提出了一种将基于规则的对感兴趣区域（Region of Interest，ROI）进行初提取的方法与最小二乘支持向量机（Least Squares Support Vector Machine，LSSVM）方法相结合的肺部结节CT图像识别方法。通过对图像中的提取结节的图像特征分析和对肺部结节区域的先验知识的分析，首先使用基于特定规则的样本提取方法，排除了一定数量的非结节感兴趣区域，然后使用基于最小二乘支持向量机的方法对其识别，从而实现了样本由低维空间到高维空间的映射，有效的避免了人工神经网络（Artificial Neural Network，ANN）等方法造成的局部最优解、高维运算复杂、泛化能力差和过学习等问题。实验结果表明，合理的选取图像特征后，使用最小二乘支持向量机的方法可以在少量增加漏检数的情况下，降低误判数。更多还原

【Abstract】 The segmentation and recognition of lung nodules in computed tomography (CT)images based on the support vector machine is the core of the computer-aideddiagnosis (CAD).The CAD can provide concise and accurate information andimprove the efficiency of medical diagnosis. The accurate image segmentationmethod, the reasonable image feature extraction and the powerful generalizationability of the machine learning are the key of the computer-aided diagnosis. The mainpoint of the segmentation and recognition of images is introduced as follows:1. In the segmentation and extraction of lung nodules aspects, the characteristicsanalysis of lung nodules section is the key in research. According to the features of theimage and the lung nodules, the highly targeted methods are used to extract thesolitary lung nodules and the adhesion lung wall of lung nodules respectively.2. In the tumor recognition aspects, the research focuses on how to reasonablychoose the appropriate image features. By the statistical theory, the identification andclassification system which expected risky is as small as possible is designed.For the above problems, the main research is as follows:1. According to the characteristics of the computer tomography image and lungnodules images, a method of image segmentation based on the improved rolling ballis used to extract the adhesion lung wall of lung nodules region; the region growingmethod is used to extract the independent nodules region in the lung area. Theextraction results could be a study sample database for the follow-up identification.The experimental results show that the improved rolling ball method can extractnodules which adhesive the wall of lung and the methods based on region growingcan extract the independent nodules.2. The recognition method for CT images of lung nodules combined the earlyregion of interest (ROI) extraction based on the rule with the least square supportvector machine (LSSVM) is proposed. By the analysis of the characteristics of theextraction nodules and the prior knowledge of the lung nodules, the sample extractionmethod based on the specific rules is used, and the certain number of the non-nodulesregion of interest is excluded. Then the method based on the least square supportvector machine is used for recognition to achieve the mapping which is from lowdimension space to high dimension space. So the problems of the local optimalsolutions, the complex high-dimensional operation, the poor generalization ability andthe over learning caused by the artificial neural network are avoided effectively. Theexperimental results show that the least squares support vector machine could reducethe false positives after the reasonable selection of the image characters when thefalse negatives are increased in a small amount.更多还原