【作者】 王文辉；

【导师】 陈武凡；

【作者基本信息】 南方医科大学 ， 生物医学工程， 2008， 硕士

【摘要】 基于内容的图像检索(CBIR),近十年来一直是研究的热点,但一直都未有重大的突破。本文从CBIR的概述入手,概述了CBIR系统的组成和所涉及的关键技术,回顾了目前的研究进展和主流的研究方法及已完成的各种系统,同时分析了医学图像的特点,最后,本文围绕CBIR在医学中的应用,研究了以下问题:(1)基于GAUSS-MARKOV随机场模型的MR图像分割(对应于本文第二章)。图像分割是图像关键内容区域的选取及量化的基础,利于用图像的关键内容信息进行检索。高斯—马尔可夫随机场模型既利用了图像像素的灰度信息,又通过像素类别标记的Gibbs光滑先验概率引入了图像的空间信息,是能较好地分割含有噪声的图像的模型。然而,Gibbs惩罚因子的确定却一直是个难点,为获得好的分割效果,通常用多个值人工尝试。本文针对此问题,提出了一种新的、简单的、类自适应的惩罚因子,其利用后验概率来自动计算,并具有各类各向异性。再将模型利用EM-MAP算法来迭代求解。最后,将本文算法应用于MR图像的分割,实验表明该算法能自适应地、有效地分割噪声图像,并具有较高的正确分类率和类正确分类率。(2)基于弹性配准的图像感兴趣区域定位(对应于本文第三章)。疾病的诊断,主要是依据病变区域的形状,大小,及其所在的解剖位置,因而病变区域为感兴趣区域。感兴趣区域的确定有利于在CBIR中只关注医学图像信息的主要部分,得出利于检索的关键信息,使用于检索的特征量显著减少。目前,众多的CBIR都提供了ROI确定的方法,但往往是采用人工交互的方式。如Blobworld中,则先采用图像分割的方法,据待检索图像的多种性质,如纹理,色彩,极性,将待检索图像分割为多个区域,然后通过交互的方式,手工确定一个ROI。本文则尝试使用弹性配准的方法来自动的确定ROI。(3)图像特征提取(对应于本文第四章)。图像特征提取是图像检索的基础与关键。常用的特征提取方法有:用傅立叶描述符和不变矩表示的形状特征,小波纹理特征,Gabor小波纹理特征。但传统小波变换具有的两大缺点:平移变化性,缺乏方向性。Gabor小波具有方向选择性,但Gabor小波由于非正交性具有大量的数据冗余且不可逆向重构,并具有较高的计算代价。在仔细分析了以上小波变换的缺点后,本文重点介绍了能克服以上小波缺点的双树复数小波变换(DT-CWT)及其原理,并将以此作为图像特征提取的方法。(4)基于DT-CWT和K-L距离(Kullback-leibler distance)相似性测度的脑部MR图像检索(对应于本文第五章)。对各图像进行二阶DT-CWT后,每级分解得到6个方向的小波子带系数,再利用广义高斯密度(Generalized GaussianDensity)函数来逼近各子带系数的直方图,即使用最大似然估计法得出与各直方图相近的广义高斯密度函数的参数,用矩匹配法得出参数,以这12组特征值来表示该图像特征。提取出图像特征后,本文则基于最大似然规则的图像检索理论,使用基于链式规则的K-L相似性测度来进行特征匹配。最后,本文对以上研究进行了总结,并对医学CBIR研究进行了展望,讨论了未来的研究方向和目标。更多还原

【Abstract】 A content-based image retrieval system in medical applications has always been the most vivid research areas in recent ten years, but no general breakthrough has been achieved. This thesis aims to bring some highlight in this field. First, it presents a brief introduction of CBIR system and its key techniques, reviews the current research development, the main research methods and the existed CBIR system, and analyzes the features of medical image. And then, centered on the medical application of CBIR, it mainly does some research about the following issues:(1) Segmentation of brain MR images based on GAUSS-MARKOV random field model, (see chapter 2). Image segmentation lays foundation for selecting and measuring the key region of content information, which is advantageous for image retrieval to use key information. Gauss-Markov random field model is the often used image segmentation model, taking advantage of both image intensity and spatial information imposed by Gibbs prior. It can be used to effectively segment the images with high levels of noise. However it is always difficult to confirm the Gibbs penalty factor. As usual, it requires a tedious trial-and-error process. So to solve this problem, this paper defines a class-adaptive penalty factor. It is automatically estimated from the posterior probability and is anisotropic for each class. Furthermore the model iteratively gets their parameters estimation in the EM-MAP algorithm. Finally, by application of this algorithm to medical image segmentation, it is proved effective.(2) Locating region of interest of images based on elastic registration, (see chapter 3). Diagnosing diseases is usually on the basis of the form and size of pathological regions and its anatomic parts. Therefore, the pathological region is also defined as region of interest. To delimit the region of interest is beneficial to CBIR for the purpose of narrowing down information area and getting key retrieval information. By this way, the image information characters for retrieval are tremendously decreased. At present, many CBIRs are capable of offering the method of confirming ROI, but all are realized by hand. For instance, in Blobworld system, firstly, the image to be retrieved is segmented into several areas according to image’s various features, such as veins, colors and polarity, and then by the way of interaction, the ROI is determined by hand. This thesis, however, tries to adopt the method based on electric registration to automatically determine it.(3) Features extraction of images based on DT-CWT (see chapter 4). Feature extraction is critical and fundamental to CBIR. The most used methods for feature extraction are: fourier descriptors, moment invariants used to express the shape features, the wavelet and Gabor wavelet used to express the texture feature. But the traditional wavelet bears two deficiencies: shift variance and lack of directionality. Though Gabor wavelet has the excellent directionality, a typical Gabor image analysis is not only expensive to compute and noninvertible, but also yields a great deal of redundant information because of its nonorthogonality. DT-CWT can overcome these deficiencies. It is a good way to abstract image features. Therefore,this thesis analyzes its theory and will adopt it in abstracting image features.(4) Brain MR images retrieval based on DT-CWT and Kullback-leibler distance similarity measure (see chapter 5). After two stages DT-CWT, an image can get 6 sub-band coefficients in 6 directions for every stage, and then each sub-band coefficient histogram is fitted with Generalized Gaussian Density function. Based on matching moment, this thesis tries to estimate parameters under the maximized likelihood rule. By means of it, 12 pairs of parameters which are image characters are abstracted from each image. After getting the features, measuring the similarity of these features is a critical problem. Grounded on the maximized likelihood selective rule, this thesis applies the Kullback-leibler distance using the chain rule to fulfilling it.更多还原