CT影像中肺结节检测与识别方法的研究

发布时间：2018-04-29 01:28

本文选题：CT影像 + 肺结节　；参考：《电子科技大学》2017年硕士论文

【摘要】：随着我国空气质量的下降,雾霾天气的频繁出现,肺部疾病的发病率呈现扩大的趋势。随着医疗设备的进步与发展,CT影像成为常见的肺部检查工具。虽然医疗设备的更新给诊断的准确性带来了极大的提高,但另一方面,成像数量的增大给医生带来的是更大的工作量。为了解决这一矛盾,20世纪90年代开始出现计算机辅助检测和诊断相关研究和应用,相关应用也证明了计算机辅助检测和诊断在降低漏诊和误诊等方面有着不可小觑的作用。本文针对DICOM标准胸部CT序列影像,采用LIDC-IDRI影像库作为研究素材,从DICOM标准的影像解析、CT序列影像的肺实质分割、边界轮廓修补、肺结节检测提取和肺结节识别分类等步骤进行相关算法研究和实验论证。具体内容:(1)对DICOM影像进行结构分析,并解析影像。然后研究使用基于区域生长的快速肺实质分割算法对解析后影像进行分割,得到肺实质影像。(2)针对分割后影像边界出现的肺结节漏分割现象,研究边界修补相关算法,引入凸包和近邻点连接实现肺实质边界修补。对肺实质影像进行基于层厚的平均密度投影算法的研究,旨在抑制水平面影像中血管和支气管等结构的表现,从而凸显肺结节和血管、支气管间的影像结构差异。(3)针对平均密度投影后的影像分别研究基于GMM的单张和序列的肺结节检测提取算法,从影像中检测提取得到肺结节和结构类似的影像区域。(4)为了从提取到的结构中分类识别得到肺结节,本文研究并使用SVM分类器来对肺结节进行分类识别。由于肺结节在空间上表现为球形或者类球形结构,本文对提取的影像区域进行多平面重建,利用水平面、矢状面和冠状面三个平面上的影像来反映肺结节的形态结构,然后从三个平面中提取计算肺结节特征,进行特征数据的归一化处理。随后使用LIBSVM工具在MATLAB下研究使用网格搜索算法进行SVM核函数的参数寻优,使用提取到的特征数据进行分类器训练,并使用测试数据对分类器进行测试,通过肺癌辅助诊断系统评价指标对分类识别结果进行评价。最后本文将研究的算法进行实现,开发出一个计算机辅助诊断系统。通过大量的算法实验,本文算法在肺实质分割中分割成功率达99%以上,肺结节识别的特异性达92.9%,相比同类识别算法有较为显著的优势。
[Abstract]:With the decline of air quality and the frequent appearance of haze weather, the incidence of lung diseases is increasing. With the progress and development of medical equipment, CT imaging has become a common pulmonary examination tool. Although the renewal of medical equipment has greatly improved the accuracy of diagnosis, on the other hand, the increase of imaging quantity has brought more workload to doctors. In order to solve this contradiction, the research and application of computer-aided detection and diagnosis began to appear in the 1990s. The related applications also proved that computer-aided detection and diagnosis play an important role in reducing missed diagnosis and misdiagnosis. In this paper, DICOM standard chest CT images were analyzed by using LIDC-IDRI image library as research materials. Lung parenchyma segmentation and boundary contour repair were analyzed from DICOM standard CT sequence images. Lung nodule detection and extraction and identification and classification of pulmonary nodules were studied and demonstrated by experiments. The structure of DICOM image is analyzed and the image is analyzed. Then we use the fast segmentation algorithm of lung parenchyma based on region growth to segment the parsed image and get lung parenchyma image. (2) aiming at the lung nodule leakage segmentation phenomenon which appears in the edge of the segmented image, we study the algorithm of boundary repair. The pulmonary parenchyma boundary repair is realized by introducing convex hull and adjacent point connection. The study of the average density projection algorithm based on slice thickness of lung parenchyma images aims to suppress the appearance of blood vessels and bronchi in horizontal images, thereby highlighting pulmonary nodules and blood vessels. (3) for the images with average density projection, the extraction algorithms of lung nodules based on single sheet and sequence of GMM were studied, respectively. In order to classify and recognize pulmonary nodules from the extracted structures, SVM classifier is used to classify and recognize pulmonary nodules. Because the pulmonary nodules are spherical or spherical in space, the extracted image regions are reconstructed in multi-plane, and the morphologic structure of the pulmonary nodules is reflected by the images of horizontal plane, sagittal plane and coronal plane. Then the feature of pulmonary nodule is extracted from three planes, and the feature data is normalized. Then the LIBSVM tool is used to study the parameter optimization of the SVM kernel function using the grid search algorithm under MATLAB, and the extracted feature data is used to train the classifier, and the test data is used to test the classifier. The classification and identification results were evaluated by the evaluation index of lung cancer auxiliary diagnosis system. Finally, the algorithm is implemented and a computer aided diagnosis system is developed. Through a large number of algorithm experiments, the success rate of this algorithm in lung parenchyma segmentation is over 99%, and the specificity of lung nodule recognition is 92.99.This algorithm has a significant advantage over the similar recognition algorithm.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R734.2;R730.44;TP391.41

【参考文献】