肺部磨玻璃密度影的分割算法研究

发布时间：2018-03-18 23:40

本文选题：磨玻璃密度影　切入点：肺实质分割　出处：《哈尔滨工业大学》2016年硕士论文　论文类型：学位论文

【摘要】：肺癌是严重威胁广大人民群众的重大疾病。仅2012年,肺癌就夺走了109.8万男性,49万女性的生命。肺癌的早期确诊是延长肺癌患者五年生存率的有效途径,同时也是减轻患者家庭负担的最有效办法。磨玻璃密度影(Ground Glass Opacity,GGO)是肺癌在CT影像中的重要早期表现,但同时也是一种非特异性表现。目前,GGO的良恶性判断主要是通过检测和分析GGO的形态学特征(如大小、形状、边缘、粘连结构、是否内含实质性成分等)及特征的生长变化来进行,GGO的准确分割是以上工作的重要前提。当前,已经有部分GGO分割算法被提出,但是其分割效果均不太理想。尤其是当前的分割算法主要以分割出GGO的主体部分为目标,忽略了GGO的边缘结构(如毛刺、条索等),不仅造成了目标对象的过分割,影响了GGO面积的测量,而且会造成后期GGO特征分析的偏差、GGO良恶性判断的错误。本文提出了基于区域自适应权重的马尔科夫随机场模型的GGO分割算法,实现了GGO较为准确的分割。但该模型中的正常肺实质灰度分布模型参数严重依赖于训练样本,GGO的灰度分布模型参数需要依据GGO的初分割结果进行更新,两者都会影响到实验的运算速度和分割结果准确性。为此,本文进一步提出了基于马尔科夫随机场的二维大津法模型和混合类泊松模型来实现GGO的分割。本文内容分为以下四个部分:1、肺实质的自动分割;2、基于区域自适应权重的马尔科夫随机场(Markov Random Field,MRF)模型;3、基于MRF的二维大津法模型;4、基于MRF的有限混合类泊松模型。肺实质的准确分割能够有效的排除胸部CT图像内与GGO分割、特征分析无关因素(如CT检查床、衣物、胸壁、纵膈等)的影响,减少计算量和计算误差,提高GGO的分割准确性。本文在肺实质的分割过程中首先使用基于OSTU(大津法)的阈值法确定分割阈值、利用连通域标记法填充肺实质区域内空洞、借助投影法确定左右肺实质是否未分离及ROI区域;然后基于滚球法和凸包法来实现肺实质的边缘平滑修补。从而得到较为完整的肺实质分割结果。基于区域自适应权重的MRF模型是指该马尔科夫随机场中的转移概率是一个基于局部隶属度的自适应参数。该参数在隶属于GGO的隶属度与隶属于正常肺实质区域的隶属度有较大差别的局部区域内取得较大的值,在隶属于GGO的隶属度与隶属于正常肺实质的隶属度较为接近的区域内取较小的值。基于该模型,GGO得到了较为准确的分割。基于MRF的二维大津法是在二维大津法的类间方差计算公式中引入了基于MRF的平滑能量。该能量能够有效的调节使用最优阈值得到的GGO潜在区域的个数及潜在区域的大小,有效的抑制过小肺纹理的影响、减少后期GGO识别过程的工作量。基于MRF的有限混合类泊松模型是在考虑像素点的空间关系的前提下,使用有限个类泊松模型对肺实质的灰度分布进行拟合的一个算法模型。它充分利用了正常肺实质、GGO、高密度肺纹理等区域的灰度分布类似于泊松分布(或偏正态分布)的性质,能够更好的对像素点进行归类。本文通过专家打分法来评价基于该三种算法的GGO分割结果的优劣,通过计算分割结果的误差率、面积和形状的测量精度来分别对比评价该算法与同类算法的优劣。实验表明基于以上三种方法,都取得了较同类算法更好的GGO分割结果。
[Abstract]:Lung cancer is a major disease that threatens the people. Only in 2012, lung cancer took 1 million 98 thousand men and 490 thousand women's life. The early diagnosis of lung cancer is a effective way to prolong the five year survival rate of patients with lung cancer, the most effective way is to reduce the family burden of patients. Ground glass opacity (Ground Glass, Opacity, GGO) is an important early manifestation of lung cancer in the CT image, but also a nonspecific performance. At present, the judgment of malignant and benign GGO is mainly through the morphological feature detection and analysis of GGO (such as size, shape, edge, adhesion structure, whether contains substantive components) growth characters and changes to. Accurate segmentation of GGO is an important prerequisite for the above work. At present, there have been some GGO segmentation algorithms have been proposed, but the segmentation results are not ideal. Especially the current segmentation algorithms to segment the main GGO The body part is the goal, ignore the edge structure of GGO (such as burr, cable, etc.) not only caused the over segmentation of the target object, affect the measurement of GGO area, but also cause the deviation analysis of late GGO features, GGO diagnosis error. This paper proposes MRF model GGO segmentation algorithm based on Markov region adaptive weights, to achieve the GGO accurate segmentation. But the model of normal lung parenchyma gray distribution of model parameters depends heavily on training samples, parameters of gray distribution model according to the GGO GGO at the beginning of the segmentation results in the update, operation speed and accuracy of the segmentation results will affect the experiment. Therefore in this paper, further put forward the Markov random field model and two-dimensional Otsu method mixed Poisson model to achieve segmentation based on GGO. This paper is divided into the following four parts: 1, automatic lung parenchyma 2, regional segmentation; Markov adaptive weighted random field (Markov Random Field, 3, MRF) model; two-dimensional Otsu method model based on MRF; 4, MRF finite mixture Poisson model based on the accurate segmentation of lung parenchyma segmentation and GGO can remove the chest CT images effectively, analysis of the characteristics of independent factors (check the bed clothes, such as CT, chest wall, mediastinum etc.) effect, reduce the computation error, improve the segmentation accuracy of GGO. Based on the segmentation of lung parenchyma in the first use of OSTU based (Otsu) to determine the threshold of the threshold method, the connected domain labeling filling lung parenchyma area empty, according to the projection method to determine the left and right lung parenchyma is not isolated and ROI region; then based on the rolling ball method and convex hull method to achieve smooth edge repair of lung parenchyma. In order to get a more complete lung segmentation results. Based on the regional adaptive weight MRF model refers to the transition probability of Markov field is a local adaptive parameter based on degree of membership. To achieve greater value of local area of the GGO in which the parameters belong to membership membership and belongs to the normal lung parenchyma area has a larger difference in the smaller regional membership degree belongs to GGO the degree of belonging to normal lung parenchyma is close to the internal value. Based on this model, GGO obtained a more accurate segmentation. The two-dimensional Otsu method based on MRF is the variance in two-dimensional Otsu method between class calculation formula was introduced in MRF. Based on the number of smooth energy and potential energy of the area effectively the regulation of GGO using the optimal threshold potential area to the size of the effective inhibition effect of small pulmonary veins, reduce the late GGO recognition process work. MRF finite mixture model based on Poisson is considered as In the premise of spatial relations, an algorithm of gray model of lung parenchyma distribution was fitted using finite Poisson model. It makes full use of the normal lung parenchyma, GGO, high density lung texture regions gray distribution similar to the Poisson distribution (or partial normal distribution) properties can be better the classification of pixels. In this paper, by using expert scoring method to evaluate the segmentation results of three algorithms based on the merits of GGO, through the calculation results of segmentation error rate, accuracy of size and shape were compared to evaluate the algorithm with the similar algorithm. Experimental results show the advantages and disadvantages of the above three methods are made based on compared to other algorithms better GGO segmentation.

【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：R734.2;TP391.41

【相似文献】