【作者】 周生俊；

【导师】 程远志；

【作者基本信息】 哈尔滨工业大学 ， 机械电子工程， 2014， 博士

【摘要】 医学图像分割是医学图像处理与分析的基础，该问题的解决不仅直接影响到计算机图形图像技术在医学中成功应用而且有重要的理论和实际意义。医学图像分割是一个提取感兴趣区域的过程，其分割结果可以为随后的疾病诊断、治疗方案规划以及治疗效果评估等提供参考。CT由于具有较高分辨力，能更清晰的彰显解剖结构和病变组织等特点，使其广泛地应用到许多系统的疾病诊断。因此，研究图像分割方法在CT图像中的应用具有非常重要的意义。本文主要针对髋关节、包含胸膜结节的肺部以及肝脏等CT图像进行研究，其目的是构建准确、自动的分割方法为医生的诊断和治疗提供帮助。正常的髋关节由股骨头和髋臼两部分组成。在髋关节CT图像中，由于股骨头和髋臼之间连接区域非常狭窄以及一些疾病等原因导致骨密度不均匀，使得准确分割髋关节CT图像变得非常困难。针对上述问题，本文提出一种基于迭代自适应阈值分类和贝叶斯判别分析技术相结合的三维髋关节CT图像自动分割方法。该方法首先利用形态学增强技术提高连接区域与骨组织之间的灰度对比度，在随后过程中，针对阈值分割结果使用基于贝叶斯判别方法的迭代自适应分类最终实现股骨头和髋臼之间的分离。由于在上述分割方法中多次使用数学形态学，使得分割出的结果往往被过分光滑而丢失骨边缘的细节；再者，形态学操作依赖于结构元素的选取，选择不同形状或者尺寸的结构元素会对分割结果带来一定的影响。针对上述问题，本文提出一种基于骨曲面顶点法线方向灰度变化的边缘校正算法。通过该校正算法不仅可以实现骨边界体素的准确定位，同时可以得到髋关节的三维可视化结果。实验结果验证了本算法的准确性以及临床适用性。由于肺部充满空气，相比于周围组织具有较低密度，因此阈值方法是一种常用的肺部CT图像分割方法。但对于包含胸膜结节的肺部CT图像来说，由于胸膜结节位置以及大小的多变性且与周围组织具有相似的密度，阈值方法难以准确的将其包含；再者靠近纵膈区域的高密度的肺部血管也被阈值方法排除在外导致肺门区域的凸凹不平，传统方法通常采用形态学方法进行光滑，但是形态学过分依赖结构元素的选取。针对上述问题，本文提出一种准确自动的肺部CT图像分割方法和一种有效的肺部边界校正和光滑算法。该方法利用模糊C均值实现肺部的快速自动分割。对于包含胸膜结节或者肺部血管的切片，本文提出一种基于迭代权重平均和自适应曲率阈值相结合的边界修复和光滑算法。该方法可以自动准确检测胸膜结节和肺部血管区域并将其光滑包含在肺部分割结果中。实验结果验证了本算法的快速性以及有效性。在肝脏CT图像中，由于腹部器官之间的低对比度、器官病变的存在以及个体之间器官形状的差异，使得传统的仅依赖图像灰度信息的分割方法难以取得较好的肝脏分割结果，传统方法往往容易导致肝脏分割结果的泄露。针对上述问题，本文提出一种基于对比增强CT图像的三维肝脏分割方法。分割方法由训练相和测试相两部分组成。在训练相中，利用主成分分析方法形状训练得到肝脏平均形状强度模型以及在各主成分上的形状变动。对测试集中的每一个目标图像，首先通过与所有图谱进行相似度匹配得到该目标图像的最有可能肝脏区域，随后利用最大化后验概率分类概率地图以及在最有可能肝脏区域应用基于窄带技术的形状强度先验水平集演化方法得到肝脏准确分割结果。实验部分验证了本算法的准确性。更多还原

【Abstract】 Medical image segmentation is the basis of medical image processing. Solving this problem not only directly affect the successful application of computer graphics technology in medicine but also has important theoretical and practical significance. Medical image segmentation is a process to extract the region of interest and segmentation result can provide a reference for subsequent diagnosis, designment of treatment programms and evaluation of treatment regimens. Due to its higher resolution, sharper highlight anatomical structures and lesions, CT has been widely applied to many systems for disease diagnosis. Therefore, the study of image segmentation method in CT images has very important significance. This paper studies segmentation methods in the hip joint, lungs with juxtapleural nodules and liver CT images, resectively. The aim of this study is to build accurate, automatic segmentation methods to assist the doctor’s diagnosis and treatment.A normal hip joint is composed of two parts: femoral head and acetabulum. In hip joint CT images, due to the narrow inter-bone space of the connecting area between femoral head and acetabulum and the uneven bone density caused by diseases, it is difficult to segment the femoral head and the acetabulum accurately. In response to these problems, this paper proposes an automatic segmentation method for the hip joint from three-dimentional volumetric CT images, which combines iterative adaptive threshold classification and Bayes discriminant analysis techniques. Our method uses morphological enhancement techniques to highlight the intensity contrast between the connection area and the bone tissue. In the subsequent process, for the segmented results obtained by thresholding, we use iterative adaptive reclassification scheme that is based on Bayes discriminant analysis to achieve the separation of the femoral head and the acetabulum.Since the above segmentation method repeatly uses mathematical morphology, this makes the segmented results lose bone details caused by overly smoothing. Moreover, morphological operations depend on the selection of structural elements. Choosing a different shapes or sizes of structural elements will take some influence on segmentation results. In response to these problems, a boundary correction algorithm based on gray changes in the vertex normal direction of bone surface is presented. By the correction algorithm, we can not only locate the bone boundary voxels accurately but also can get three-dimensional visualization results of objects to be segmented. Experimental results demonstrate the accuracy and clinical feasibility of the proposed algorithm.As lungs fill with air and have a lower density compared with the surrounding tissue, thresholding is a common method used for lung segmentation. However, for CT images including juxtapleural nodules, due to the variability of the position and size of the juxtapleural nodules and the similarity intensity with surrounding tissue, it is difficult for thresholding to include these juxtapleural nodules accurately. Furthermore, high density pulmonary vessels are also ruled out from the lung area, which bring about indentations and salience in the lung boundary near the mediastinum. Traditionally, mathematical morphology is generally used for smoothing the lung boundary. However, morphology highly relies on the selected structure element. To cope with these problems, we develop an accurate and fully automatic method for segmenting lung boundary in chest CT images and an efficient scheme for smoothing and correcting the segmented lung boundary. The proposed method uses fuzzy c means algorithm to achieve the fast lung segmentation. For the slices containing juxtapleural nodules and pulmonary vessels, a contour correction and smoothing algorithm is proposed which is based on iterative weighted averaging and adaptive curvature threshold. This method can be automatically and accurately detect juxtapleural nodules and pulmonary vessels and smoothly include in the lung segmentation. Experimental results demonstrate the rapidity and effectiveness of the algorithm.In liver CT images, due to the low contrast of adjacent organs, the presence of abnormalities and the highly varying shapes between subjects, it is difficult for traditional segmentation methods only relying on gray intensity information to obtain good segmentation results, which often lead to the leakage of liver. In response to these problems, this paper proposes a three-dimensional liver segmentation method from contrast enhancement CT images. The proposed segmentation method contains a training and test phrase. In the training phase, we use principal component analysis method to get the shape-intensity model of the liver as well as the variabilities of shape and intensity on different modes. For each target image in the test phase, the most likely liver region is first obtained by computing the similarities between the atlases and the target image. Then, the precise liver segmentation result is obtained by a maximum a posteriori classification of probability map followed by applying a shape-intensity prior level set segmentation implemented by narrowband technique inside the most likely liver region. Experimental results demonstrate the accuracy of the algorithm.更多还原