CT图像的肝脏及肺部肿瘤检测方法研究

发布时间：2018-02-06 03:25

本文关键词： 概率图谱水平集可变环形滤波器自适应模糊聚类分割与检测　出处：《哈尔滨工业大学》2016年博士论文　论文类型：学位论文

【摘要】：计算机辅助分割与检测在人体重要器官的早期诊断与治疗、外科规划与导航等方面发挥着重要的作用。CT医学成像技术,因其高分辨率、低成本等特性,被广泛应用于人体重要器官的医疗诊断。然而,人体重要的腺体器官——肝脏,其CT成像由于对比度较低、形状不规则、灰度不均匀、相邻组织干扰等因素,往往导致计算机辅助分割与检测非常困难;人体重要的呼吸器官——肺,其CT成像虽然对比度较高、但由于血管、支气管、以及胸膜附近异常的存在,往往也对其分割与检测的准确性造成很大影响。针对这些问题,本文以肝脏和肺的精准分割及其肿瘤检测为研究目标,在高分辨率螺旋CT图像中,研究肝脏的分割方法、肝脏肿瘤的检测技术;肺的分割方法、肺肿瘤的检测技术。研究的目的是,通过本文的研究,提升这两大重要器官自动分割、自动检测的精准性和自动化程度,帮助放射科医生提高诊断效率,降低肝癌与肺癌致死风险。针对肝脏分割过程中,腹部多个器官相互邻近、对比度低以及肿瘤异常导致的分割难题,本文提出了一种基于概率图谱与水平集混合技术的自动分割方法。首先,通过基于特定病人的肝脏概率图谱与测试图像之间的相似度,生成最大可能性的肝脏区域;然后,根据直方图分析将最大可能肝脏区域进行后验概率分类,通过最大后验概率映射实现肝脏的粗分割;最后,通过基于形状-灰度先验模型的水平集演化技术,对粗分割轮廓优化,并获取最终细分割的肝脏实质。该新方法是首次应用于肝脏分割,实验表明该新算法具有很高的自动分割精度,且能够有效处理器官邻近以及包含肿瘤的肝脏分割情形。肝脏分割的复杂性,以及肝脏内异常的多样性,是肝脏肿瘤自动检测技术发展的主要障碍。本文在前期肝脏自动分割研究的基础上,针对肝脏肿瘤的近球形特性,提出了基于可变环形滤波器的自动检测方法。首先,基于图谱与水平集混合技术获得肝脏分割轮廓;然后,利用可变环形滤波器的半径自适应特性检测可疑肿瘤区域,并提出了基于灰度权重的距离转换函数,来对滤波后的种子点进行对比度增强;最后,利用支持向量机分类算法对候选肿瘤进行基于特征的分类,获取最终肿瘤的位置与类别信息。这是可变环形滤波器首次应用于肝脏肿瘤检测,实验表明,该方法对于近球形的不同尺寸的肝脏肿瘤具有较高的检测敏感度。针对肺分割的过程中,附胸膜肿瘤容易造成的过分割问题,本文提出了基于边界重构与凹形区域修正的肺分割技术。首先,借助形态学滤波与连通区域分析技术获取胸部区域掩码,完成胸部提取;然后,利用基于对角线的边界跟踪技术,实现肺部轮廓的初始分割;接着,基于最大代价路径实现左右肺分离,并应用基于弧长重采样的边界重构技术,实现肺轮廓的锯齿状边缘平滑;最后,应用提出的凹凸判断函数对附胸膜凹形区域进行检测与修复,通过降低由附胸膜肿瘤造成的过分割,提高肺实质的分割精度。实验表明,该方法在保持分割精度与复杂度的同时,能够有效降低附胸膜肿瘤造成的过分割误差。肺分割技术的相对成熟,促进了肺肿瘤自动检测技术的快速发展,很多学者提出了基于不同分类器的检测方法。然而,基于分类器的肺肿瘤自动检测技术,在检测精准度、假阳性、系统运行速度等方面的性能仍然不够高,无法直接应用于临床。本文在前期肺分割的研究基础上,提出一种基于改进的模糊C均值聚类技术的肺肿瘤自动检测方法。首先,利用可变环形滤波器将可疑的近球形的候选肿瘤提取出来,并利用球度阈值对近似管状结构进行筛除,初步降低假阳性。然后,对于候选肺肿瘤实施特征选择,特征计算,应用改进的自适应模糊聚类技术对训练集、测试集分别进行聚类。最后,设计马哈拉诺比斯距离函数,判断测试集与已标记的训练集之间的相似性,从而进一步降低假阳性,获取最终的肺肿瘤分类信息。该方法能够有效提高分类器参数的自适应性,并且在有限的特征训练的条件下,获得与主流算法相近的分类水平。总之,本文提出了针对肝脏与肺的分割及其肿瘤检测的自动化方法,并采用医院提供的数据进行了实验与分析,其中部分算法参加了国际竞赛,整体上取得了预期的效果,证明了方法在计算机辅助分割与检测中的应用前景。
[Abstract]:Early diagnosis and treatment of computer aided segmentation and detection in human organs, surgical planning and navigation plays an important role in the.CT medical imaging technology, because of its high resolution, low cost and other characteristics, are widely used in medical diagnosis of important human organs. However, the human body important organs - liver, the CT imaging because of low contrast, irregular shape, uneven gray, tissue adjacent interference and other factors, often lead to computer aided segmentation and detection is very difficult; human respiratory organs: lung, the CT as though the high contrast, but because the blood vessels, bronchi, and pleural abnormalities are often near, the segmentation and detection accuracy is greatly affected. To solve these problems, based on the liver and lung cancer detection and accurate segmentation as the research object, in high resolution spiral CT images And the segmentation of the liver, liver cancer detection technology; lung segmentation method, detection of lung cancer. The purpose of the study is that through this study, to enhance the two important organs of automatic segmentation, automatic detection accuracy and automation degree, help doctors to improve the efficiency of diagnosis, reduce liver cancer and lung cancer the risk of death. For liver segmentation, multiple abdominal organs adjacent to each other, the problem of low contrast and tumor segmentation caused by abnormal, this paper proposes a probability map and level set based on the hybrid technique of automatic segmentation method. Firstly, the similarity between the probability map and liver test image specific patient based on liver region to generate the maximum possible; then, according to analyzing the histogram of the maximum possible liver area of posterior probability Classification, the maximum a posteriori probability mapping to achieve coarse liver Finally, through the cut; gray shape prior model based on the level set evolution technology, the coarse segmentation contour optimization, and obtain the final segmentation of the liver parenchyma. The new method is first applied to the liver segmentation experiments show that the new algorithm has the automatic segmentation of high precision, and can effectively and adjacent liver processor officer contains the tumor segmentation. The complexity of liver segmentation, and liver abnormality in diversity is a major obstacle to technical development of automatic detection of liver tumor. Based on automatic segmentation in the early stage of liver, nearly spherical according to the characteristics of liver tumors, the automatic detection method based on variable ring filter. Firstly, Atlas and the level set technique to obtain the mixed liver segmentation based on the contour; then, using regional characteristics to detect suspicious tumor with variable radius adaptive loop filter, and based on gray The weights of the distance transform function of the seed point after filtering contrast enhancement; finally, using support vector machine classification algorithm for candidate classification based on tumor characteristics, tumor location and obtain the final category information. This is a variable loop filter is first applied to the detection of liver tumors, experiments show that this method has good detection for liver tumors of different sizes of nearly spherical sensitivity. In the process of lung segmentation, with pleural tumor easy to cause the over segmentation problem, this paper presents the boundary reconstruction and concave region modified lung segmentation based on technology. Firstly, by means of morphological filtering and connected region analysis techniques to obtain the chest region mask, complete extraction of the chest then, using the diagonal; boundary tracking technique based on the realization of initial lung contour segmentation; then, the maximum cost path based on the separation of left and right lung, and The application of boundary reconstruction technology of arc length based on resampling, realize the jagged edge smoothing lung contour; finally, put forward the application of convex judgment function for detecting and repairing of pleural concave region, by reducing an attached pleural tumor caused by the segmentation of lung parenchyma, improve the segmentation accuracy. Experimental results show that this method in keep the segmentation accuracy and complexity at the same time, can effectively reduce of pleural tumor caused by the segmentation error. Lung segmentation technology is relatively mature, and promote the rapid development of automatic detection of lung cancer, many scholars have proposed different classifiers based on detection method. However, the automatic detection technology based on classifier in the detection of lung cancer. The accuracy, false positive, the performance aspect of the system operation speed is still not high enough, can not be directly applied to clinical. Based on the study of early lung segmentation, is proposed based on an improved Automatic detection method of fuzzy C mean clustering technique of lung tumors. First, using the variable loop filter will be nearly spherical suspicious candidate tumor is extracted, and the approximate tubular structures were screening by ball threshold, primarily to reduce false positive. Then, the candidate for lung cancer the implementation of feature selection, feature calculation, fuzzy clustering technology the training set using the improved adaptive clustering, respectively test set. Finally, the design of Mahalanobis distance function, the similarity between the test set and the labeled training set, so as to further reduce the false positive, obtain the lung tumor classification information in the end. The adaptive method can effectively improve the classifier parameters, and the characteristics of training under the condition of limited access, similar to the mainstream level of classification algorithm. In short, this paper presents a segmentation and detection of liver tumor and lung self Experiments and analysis were carried out with data provided by hospitals. Some of them participated in the international competition, and achieved the expected results as a whole. It proved the application prospect of the method in computer-aided segmentation and detection.

【学位授予单位】：哈尔滨工业大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：R730.44;TP391.41

【相似文献】