基于多约束图割的肺4D-CT图像肿瘤分割算法研究

发布时间：2018-03-17 16:09

本文选题：肺4D-CT　切入点：图像分割　出处：《南方医科大学》2017年硕士论文　论文类型：学位论文

【摘要】：肺癌是常见的恶性肿瘤之一,严重威胁人类健康,而放射治疗是治疗肺癌的有效手段。肺4D-CT在肺癌的放射治疗中扮演着重要角色,它能真实地反映人体各器官组织和肿瘤随呼吸运动的变化规律,这些信息对于肺癌的精确放疗意义重大。利用肺4D-CT图像可以针对患者的运动特征制定个体化的精确放疗计划,降低靶区外放边界,提高靶区受照射剂量,减少周围正常器官组织所受的照射剂量。但是一套肺4D-CT数据是由多个相位的图像组成,通常包含上千张图片,医生勾画靶区的工作量就成倍增长,这会影响医生对靶区的准确判断。而在放疗过程中靶区定位十分关键,其准确性直接影响放疗效果,因此肺4D-CT的靶区勾画是一个耗时耗力的工作。利用图像分割技术来帮助医生勾画靶区就是有效的解决办法。图像分割是根据人们的需要,将图像划分成各具特性的不同区域。医学图像分割就是将感兴趣的器官或组织提取出来。本文就是对肺4D-CT图像进行分割,提取出各个相位上的肿瘤图像,帮助医生确定靶区。虽然近年来图像分割算法层出不穷,但是CT图像上肿瘤与周围组织的对比度较低,而且肺肿瘤大小形态位置各异,容易侵犯周围正常组织器官而发生粘连,边界模糊,准确分割肺肿瘤图像依然十分富有挑战性。而且从降低医生工作负担的角度来说,对分割算法的自动化程度也要求较高。本文针对上述问题,提出了两种基于不同约束图割的肺4D-CT图像分割算法,是在图割算法的基础上对其进行了改进,提高了它的分割精度和自动化程度,协助医生勾画靶区,降低他们的工作量,使4D-CT更方便地应用于肺肿瘤精确放疗。其一,本文提出了一种基于星形先验和图割的肺4D-CT肿瘤自动分割方法。首先在4D-CT初始相位图像上,由医生选取目标种子点,以此种子点为中心,形成一个N×N×N大小的初始目标块,将肿瘤全部包含在其中。N的大小由医生观测肿瘤大小估计得到。然后采用运动估计中的完全搜索块匹配算法,获得下一相位图像中与初始目标块最相似的目标块,同时估计出它们之间的运动位移,以此类推,可以得到所有相位的目标块以及对应块之间的运动位移。接下来利用这些运动位移和初始相位目标种子点的位置,计算出其余各相位目标种子点,作为星形先验的中心点。最后在各相位的目标块上使用结合星形先验的图割算法,即可得到肿瘤分割结果。实验结果表明,此方法的分割准确性优于传统图割算法,同时也提升了算法的自动化程度。其二,本文提出了一种基于图割的利用肺4D-CT上下文信息的多相位肿瘤联合分割方法。该方法是将肺4D-CT各相位图像联合构建成一个全局网络图,每个相位的图像都是一个子网络图,上下文信息作为约束项加入到各个子网络图之间,也就是给相邻网络图的对应节点之间加上边。在多相位网络图的基础上,我们构建了新的全局能量函数,包含各相位的区域项、边界项和新增加的上下文信息约束项。其中区域项和边界项的构造与原始图割算法相同,而上下文信息约束项参考了 Potts模型,用来惩罚相邻相位对应体素点标号不一致的情况。通过优化新的能量函数就能完成多相位肿瘤的联合自动分割。该方法进一步减少了用户交互,用户只需在某一相位图像上选取目标和背景种子点,所有相位上的肿瘤都能被自动分割出来。我们在十套肺4D-CT数据上进行了实验,从视觉和量化结果来看,该方法的分割结果优于未加入上下文约束的图割算法以及结合星型先验的图割算法。
[Abstract]:Lung cancer is one of the most common malignant tumor, a serious threat to human health, and radiation therapy is an effective therapy for lung cancer. Lung radiotherapy plays an important role in 4D-CT lung cancer, it can reflect the human organs and tumor tissue changes with the respiratory motion, precise radiotherapy for lung cancer and the significance of these information. The plan of precise radiotherapy for lung 4D-CT images can be individualized according to the motion characteristics of the patients, reduce the target area on the outside of the boundary, improve the radiation dose target area, reduce the radiation dose to surrounding normal organs. But by a set of lung 4D-CT data is composed of a plurality of phase images, usually contains thousands of pictures the doctor, target delineation of the workload doubled, it will affect the doctor to target accurately. In the course of radiotherapy target positioning is very important, its accuracy directly affects the radiotherapy Therefore, the target delineation of lung 4D-CT is a time-consuming work. Segmentation technology to help doctors target delineation is an effective solution using image. Image segmentation is according to the needs of the people, will be divided into different areas of the image. The characteristics of the medical image segmentation is of interest to the organ or tissue this article is extracted. The segmentation of lung 4D-CT images, extract tumor image of each phase, help the doctor determine the target area. Although in recent years, the image segmentation algorithm of CT image but emerge in an endless stream, the tumor and surrounding tissue of low contrast and lung tumor size and shape at different locations, easy invasion of surrounding normal organs and tissues adhesion, fuzzy boundaries, accurate segmentation of lung tumor image is still very challenging. But from the perspective of reducing the doctors work burden, the degree of automation of the segmentation algorithm Also higher. Aiming at these problems, put forward two kinds of segmentation algorithms based on different constraint graph cut lung 4D-CT images, is based on the graph cut algorithm on its improvement, improve its segmentation accuracy and automation, to assist doctors to target delineation, reduce their workload, make the 4D-CT more convenient application in lung tumor precise radiotherapy. First, this paper presents a method for automatic segmentation of Hoshi Gata's transcendental and graph cut based on 4D-CT lung cancer. First in the initial phase of 4D-CT image, by the doctor to select the target seed point, the seed point as the center, to form the initial target block of a N * N * N size all included in the tumor, which the size of the.N by the observation of tumor size. Then the doctor estimated full search motion estimation block matching algorithm, to obtain the next phase in the image and the initial target block is most similar to the target block, at the same time The estimated motion between them and so on, can get all of the target block and the corresponding block phase between the displacement. Then using the displacement and the initial phase of target seed location, calculate the rest of the phase of target seed point as the center point of star shape prior. The last target block in each phase of combined with the use of a priori star graph cut algorithm, the segmentation results of tumor can be obtained. The experimental results show that the segmentation accuracy of this method is better than the traditional graph cut algorithm, but also enhance the degree of automation of the algorithm. Secondly, this paper proposes a segmentation method combined with multi phase 4D-CT lung tumor using context information based on the graph cut. The 4D-CT method is the pulmonary phase image are combined to construct into a global network map, the image of each phase is a sub network, context information as a constraint and Into each sub network diagram, is to add on the corresponding node between adjacent network diagram. Based on multi phase network graph, we construct the global new energy function, including the phase boundary region, and additional context information. The regional tectonic constraints and boundary the same with the original graph cut algorithm, and context information constraints refer to the Potts model, to punish the adjacent phase corresponding voxel labeling isinconsistent. Combined with automatic segmentation of multi phase tumor by optimizing energy function. The new method can further reduce the user interaction, the user only need to select the object and background seeds in a phase image, all phase of the tumor can be automatically separated. We conducted experiments on ten sets of lung 4D-CT data, from the visual and quantitative results, the method of The cutting results are better than that of the graph cut algorithm that does not join the context constraint and the graph cutting algorithm combined with the star type prior.

【学位授予单位】：南方医科大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R734.2;TP391.41

【参考文献】