【作者】 程丹松；

【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， 2010， 博士

【摘要】 医学图像分割技术是医学图像处理与分析领域的重要研究课题之一,其目的是将图像中具有某些特殊含义的区域分割出来,同时提取相关的特征数据,从而为临床诊疗和病理学研究提供可靠依据。医学图像分割有着一般图像分割的共性问题,同时,由于人体解剖结构的复杂性、组织器官形状的不规则性及个体之间的差异性,一般图像分割方法直接应用于医学图像并不能得到理想分割效果,为此寻求有效的医学图像分割方法一直是备受关注的研究热点。图像分割的目标是将特征域内特征值相等或相近的像素作为同质区分割开来,同质区内部的像素在特征域内与其相邻近的像素具有很高的一致性。因此,考察像素的邻域状态是图像分割的重要手段之一。本文利用像素邻域特征实现图像分割。与以往的邻域不同,本文的邻域特征不仅仅包括邻域内像素的特征值分布,还将邻域内像素的空间分布特征进行建模,作为特征量参与分割。利用所构建的邻域特征,本论文对医学图像分割技术的关键算法和相关问题进行了研究,主要包括图像去噪、图像增强和图像分割等算法。具体包括以下几方面的工作:(1)医学图像模糊增强针对医学图像对比度较低、边缘模糊等特点,本文提出一种基于模糊理论的医学图像增强算法。首先在通过非线性变换算子对图像进行归一化的同时实现边界区域对比度拉伸,然后在变换域内采用幂次变换对图像进行进一步的对比度增强,并利用邻域信息控制增强力度,在保护同质区明显的纹理特征的同时增大区域之间的对比度。接着利用图像统计特性对多层次图像进行模糊划分,通过对各模糊子集的增强处理将算法推广为多层次模糊增强,在保护图像主要纹理特征的基础上提高不同灰度级区域间的对比度。通过与经典方法进行比较,实验结果显示本文的增强算法能很好的提高医学图像的对比度,显著提高医生临床诊断的有效性。(2)灰度医学图像分割传统的PCNN方法对噪声有很强的鲁棒性,但该方法分割效果对参数有很强的依赖性,参数选择不当,将导致欠分割或过分割,本文针对传统的PCNN模型的不足,提出基于邻域激励脉冲耦合神经网络(NIPCNN)的图像分割方法。在本方法中,点火神经元对其邻域神经元的捕捉由被捕捉神经元的强度值及其邻域决定,这个邻域包含两方面的特征,即神经元邻域元素的强度值以及强度值高于阈值的神经元的分布情况。我们将对该神经元的邻域建模,控制神经元的内部活动,决定神经元是否点火,从而实现目标区域的精确分割。实验结果表明新的模型对参数的选择依赖性明显减小,适合对医学图像的分割要求。(3)彩色医学图像分割本文通过对彩色多普勒超声图像的特点分析,提出针对该类大背景的彩色图像的分割算法。该算法采用感兴趣区域(ROI)的初步筛选,减小分割算法的处理对象,从而显著提高算法的处理速度。本文的算法基于ε邻域一致性进行分析,易于解释和实现。算法采用色差来度量像素的邻域状态,根据ε相似邻域判定准则将像素集合分成不同的等价类,然后只考虑等价类的外围边界情况将等价类演化为同质类,对应图像中不同的颜色区域。本方法在完成颜色聚类的同时完成分割,并保证分割不存在二义性。另外,根据分析确定算法的计算复杂度近似与目标区域成线性关系,因此容易实现实时处理。更多还原

【Abstract】 Medical image segmentation technology is one of the important subjects within medical image processing and analysis research field. The main purpose of medical image segmentation is to divide the image into different regions with special signification, and extract the correlative properties at the same, which can provide the credibility gist for clinic diagnose and pathology research. Besides the common properties of the image segmentation, as the complexity of human anatomic structure, the abnormity of the tissue shape and the difference among individuals, the common image segmentation methods are not fit for the medical images. Hence, to seek effective medical image segmentation method is all through the hotspot in medical image processing.The task of the image segmentation is to divid objects in an image that are touching each other into separate objects as homogeneous areas. The local characteristics of the pixels in a homogeneous area is similar to each other. Hence, it is significative to evalue the neigbourhoods of the pixels. In this thesis the segmentation methods by means of evaluing the neigbourhoods of the pixels are discussed. Differing from the existing“neigbourhoods”, it concerns the spacial distribution of the neigbours in this neigbourhoods. By employing the neigbourhoods features proposed, the dissertation discusses the key algorithms and relative issues of medical image segmentation, involving image denoising, enhancement and segmentation. Its main contents include:(1) Medical images enhancement based on fuzzy logicA medical images enhancement algorithm based on fuzzy logic is proposed in order to improve the low contrast and blurring of the image. The nonlinear operator is applied to normalization operation to enlarge the contrast in boundary regions, followed by exponential transformation. The neighborhoods are adopted to control the enhancement in order to enhance the contrast between the regions and preserve the texture in homogenous regions. For multi-level greyscale images, they are partitioned into several fuzzy sets according to their statistical properties. The multilevel enhancement is implemented by combining enhancement on each fuzzy set. Comparing the classical methods, the results obtained using the proposed method are shown to have higher contrast, thereby better representing the anatomical structures of interrogated tissue. (2) Medical gray images segmentationPCNN model is robust to noise, but the performance of the classical pcnn is sensitive to the parameters. Unsuitable parameters will lead to deficient- segmentation or over-segmentation. A neighborhood inspiring PCNN is proposed. In the proposed model, if a neuron is captured or not depends on its intensity and neibourhood, which consists two aspects: the intensity of its neibourhood neurons and the distribution of the neurons whose intensity is higher than the threshold. The neibouthood of the neuron is modeled to control the internal activity and determine if it pulses or not. The experimental results shows that the performance of new model is less sensitive to the selection of parameters.(3) Medical color image segmentationA fast segmentation method is presented based on the analysis of the color Doplor ultrasound image to deal with the images with large background. The pending area is reduced by setting region of interest (ROI), consequently, the processing is accelerated markedly. The proposed algorithm relies on an introducedε-neighbor coherence segmentation criterion which is easy to interpret and implement. The pixels are divided into several equivalence classes according their neighbourhoods measured by difference of the color, then the equivalence classes grow into homogeneity classes by merging the outer-neighboring pixels which areε-similar. Each homogeneity class is processed as a region. The segmentation is completed by color clustering. Moreover, the method has a computational complexity nearly linear in the number of image pixels in ROI, and is wieldy for realtime application.更多还原