【作者】 刘研；

【导师】 唐降龙；

【作者基本信息】 哈尔滨工业大学 ， 人工智能与信息处理， 2013， 博士

【摘要】 乳腺癌是当今一种在全世界女性当中发病率最高的恶性肿瘤疾病。早期检查是预防乳腺癌的一个非常重要的手段。由于性价比高、无放射性、副作用小等优点，超声检查被广泛应用于乳腺癌的早期诊断中。为了提高医生对乳腺超声检查的客观性和诊断效率，计算机辅助诊断系统应运而生。目前基于乳腺超声图像的计算机辅助诊断普遍采用单帧B超图像对目标的几何特征、边界特征及纹理等特征进行提取和分类。但是，由于人体体位移动、生理变化、超声声像复杂多样、良恶性肿块在超声声像上存在不同程度的交叉和重叠，单帧的特征分析必然影响诊断的准确率。此外，单一模态的图像往往不能提供临床诊断中所需要的足够信息，因此需要将不同模态的图像信息进行融合从而进行全面和综合的分析。为了克服单帧图像和单一模态特征的片面性，本文提出利用B型超声图像中的静态特征和彩色多普勒超声图像序列中的动态特征相结合的方式对肿块的良恶性进行综合分析。其中对图像进行有效的分割、配准和描述图像序列中的运动信息是静态图像和动态图像序列特征提取的核心问题。近年来，这些技术经历了深入和广泛的研究，但是在对于乳腺超声图像的处理方面仍存在一些尚未解决的关键问题。本文针对不同模态图像的特点，对乳腺超声图像分割、乳腺超声图像配准和彩色多普勒超声图像序列的特征提取等核心问题展开了相关的研究。本文所完成的工作和主要创新点如下：(1)对基于细胞自动机原理的乳腺超声图像分割方法进行研究。乳腺超声图像中的高噪声、复杂结构、模糊边界等因素是影响图像分割效果的主要原因。为解决上述问题，本文根据细胞自动机的能量传播机制，采用能量下降策略反映图像中像素点之间的空间信息，并提出种子点比较函数和局部纹理特征比较函数分别对图像的全局信息差异和局部信息差异进行建模。在此基础上，本文采用Von Neumann邻域系统和Moore邻域系统相结合的方式作为细胞自动机的演化环境，并将自适应邻域准则应用到Moore邻域系统中用来进一步地抑制噪声的干扰。该分割算法有助于识别图像中的模糊边界，对噪声具有鲁棒性，能够在较为简单的初始条件下准确地对乳腺超声图像进行分割。(2)对基于光流场的全自动乳腺超声图像配准方法进行研究。为了克服乳腺超声图像高噪声、目标结构复杂等因素的干扰，本文将惯性原理应用到配准过程中，并产生一个能够使像素点在短时间内的运动过程中保持一个原有运动倾向的惯性力，从而克服运动过程中所受噪声点的干扰。在此基础上，本文根据牛顿第二定律思想，在每次迭代过程中利用惯性力逐步地改变合外力的大小和方向，进而融入到光流场的计算方程中对每个像素点的加速度重新进行估计。此外，本文还采用牛顿第三定律的思想对算法的收敛速度进行提高。文中算法有助于在克服噪声干扰的同时保留图像的细节，并能够快速和准确地对乳腺超声图像进行配准。(3)对基于B超图像和彩色多普勒超声图像序列的乳腺肿块分类方法进行研究。本文首先采用颜色矩、颜色信息熵等不同的统计方法对血流的形态学信息进行建模。为提取血流动力学特征，本文首先对每个图像序列下的不同的血流信号位置进行了配准。然后结合医学背景知识，采用图像分格法模拟临床诊断中的取样容积，并根据彩色多普勒成像原理，从每个窗口内产生一个离散多普勒波形信号，进而对一系列重要的血流动力学特征进行建模。此外，本文提出一个“速度聚合向量”方法自动地寻找局部血流动力学特征提取的感兴趣区域。最后，将这些特征输入到支持向量机中进行肿瘤良恶性的划分。本文提取的特征提取策略有助于提高乳腺超声计算机辅助诊断的精度，同时降低误诊率和漏诊率。更多还原

【Abstract】 Breast cancer is one of the most common cancers and affects women’s healthseriously. Early detection is an effective way to control the disease. Breastultrasound imaging has become one of the most prevalent and popular approachesfor breast cancer diagnosis due to the fact that it is radiation-free, non-invasive,painless, cost-effective and portable. In order to improve the effectiveness andaccuracy, computer-aided diagnostic (CAD) techniques are more and more appliedto clinical practice. Existing classification methods usually extract the geometricfeatures, boundary features and texture features from a single frame captured fromB-Mode ultrasound video. However, due to body position changes, physiologicalchanges, complexity and diversity of ultrasound imaging, cross and overlap withbenign and malignant tumors, analyzing the image with the features extracted fromthe single frame will affect the diagnostic accuracy inevitably. In addition, singlemode image can not provide enough clinic information; therefore, it is necessary tofuse the image with different model to achieve a comprehensive and synthesizedanalysis.To overcome the one-sidedness of single mode image, this dissertationproposed to integrate the features of B-Mode ultrasound image and color Dopplerultrasound image sequence to analyze the benign and malignant breast tumorsynthetically. Wherein, image segmentation, image registration and describing thedynamic information of image sequence are the key problems of feature extractionfrom static image and dynamic image sequence which have been well studied.However, there are still problems in processing of breast ultrasound images.According to the different image characteristic, this dissertation proposed severalmethods according to a series of conventional processes including imagesegmentation, image registration, feature extraction and classification to solve theproblems above to the breast ultrasound images.The main research work and contribution of this dissertation are:(1) A novel breast ultrasound (BUS) image segmentation algorithm based oncellular automata is studied. Due to high noise, complicate structure and blurryboundary, breast ultrasound image segmentation is a difficult task. To overcome the problems, an energy decrease strategy is used for modeling the spatial relationinformation of pixels according to the energy transition principle of cellularautomata. Then, a seed information comparison function and a texture informationcomparison function are proposed for modeling the global image informationdifference and local image information difference. In addition, two neighborhoodsystems (von Neumann and Moore neighborhood systems) are integrated as theevolution environment, and a similarity-based criterion is used for suppressing noiseand reducing computation complexity. The proposed method is helpful to handleBUS image with blurry boundaries and low contrast well, segment BUS imageaccurately and effectively within a simple initial condition.(2) A fully automatic non-rigid image registration algorithm based on opticalflow principle is studied for registration of BUS images. To overcome the affectionof speckle noise, complicate structure of BUS image, this dissertation proposed toapply the inertia principle to the image registration, and an “inertia force” derivedfrom the local motion trend of pixels in a Moore neighborhood system is producedand integrated into optical flow equation by the Newton’s second law to estimate theacceleration direction, which is helpful to handle the speckle noise and preserve thegeometric continuity of image. In addition, the proposed method integrated the ideaof Newton’s second law to accelerate the convergence speed. The proposed methodis helpful to register ultrasound images efficiently, robust to noise, quickly andautomatically.(3) A breast tumor classification method based on B-Mode ultrasound imageand color Doppler image sequence is studied. First, the color moment and colorentropy methods are utilized for modeling the vascularity features. For extractinghemodynamic features, an image registration method is utilized for mapping theposition of corresponding blood signals in different phases. Then the color Dopplerimage is divided into non-overlapping lattices. From each lattice, a discrete Dopplerwaveform is constructed from the registration results and several importanthemodynamic features are extracted. Furthermore, a velocity coherence vectormethod is proposed to design to the region of interest for extracting the localhemodynamic features. Finally, these features are employed to discriminate benignmasses from malignant masses by using the support vector machine classifier. Theproposed method is helpful to improve the true-positive and decrease the false-positive diagnostic rate, which is useful for reducing the unnecessary biopsyand death rate.更多还原