【作者】 张小峰；

【导师】 张彩明；

【作者基本信息】 山东大学 ， 计算机应用技术， 2014， 博士

【摘要】 图像分割技术是计算机视觉、图像理解等领域中最重要的研究课题之一,也是医学图像处理的核心内容和主要课题。医学图像分割是指根据图像中像素的相关特征,将图像分割为互不相交的不同组织或器官。基于医学图像分割技术,可以提取医学图像中感觉兴趣的器官/组织,对病变组织等进行定性和定量分析。同时,医学图像分割又是三维重建等影像处理的基础,分割的效果直接影响到三维模型重建的精确性。因此,研究医学图像分割技术具有非常重要的现实意义和实用价值。由于医学图像自身的复杂性,医学图像分割技术已经成为制约医学图像处理进一步应用和发展的瓶颈。考虑到医学图像中的强度不一致、部分容积效应和噪声等现象,本文拟对基于模糊聚类算法的医学图像分割技术进行研究。与其他分割方法相比,模糊聚类算法是一种典型的“软分割”方法,它允许像素以不同的隶属度同时隶属于不同的组织和器官,可以有效地解决医学图像中的部分容积效应现象。基于模糊聚类算法对医学图像进行分割时,可以从原图像中保留尽可能多的信息,因而可以获得较好的分割效果。然而,传统的模糊聚类算法由于没有考虑像素的邻域信息,对医学图像中的噪声、伪影等非常敏感,分割效果不是很理想。同时,算法在运行效率上也较为低下,无法满足医学图像处理的实时要求。针对这两个问题,在现有的改进模糊聚类算法的基础上本文重点进行了以下研究工作：(1)针对FCM算法效率低下的问题,本文从聚类中心计算的角度进行了分析,认为算法效率低下的一个重要原因在于聚类中心的计算需要涉及图像中的所有像素。基于此,本文提出了基于分层技术的图像分割技术,假设聚类中心仅由相应聚类内的像素强度值决定,与隶属于其他聚类的像素无关。算法首先基于阈值分割技术将医学图像粗略地分割为不同的组织／器官,然后基于模糊聚类算法对初略分割的图像进行修正,从而可以提高分割算法的运行效率。(2)研究了医学图像分割的实时分割技术。针对FCM算法效率低下以及相关改进算法分割效果不理想的问题,本文首先对现有的分割算法进行深入分析,认为运行效率低下的原因是算法没有充分挖掘图像中隐藏的信息。基于相关分析,提出了基于直方图的FCM算法。算法利用峰值检测技术将医学图像的直方图分割为不同的区间,并假设聚类中心取决于相应的区间,与其他区间无关。利用获得的区间在图像的直方图上基于模糊聚类算法对图像进行分割可以进一步提高算法的运行效率。一般情况下可以在0.1秒内完成医学图像的分割,满足医学图像分割的实时性要求。(3)对基于邻域信息的图像分割技术进行研究。FCM的改进算法FCMS算法在进行图像分割时,采用常数α作为邻域像素对中心像素的影响因子,这样处理可以抑制噪声的影响,然而在对不含噪声的图像进行分割时,目标的边缘部分存在模糊现象。针对该问题,本文在研究过程中基于像素的空间信息和灰度信息提出了像素的相关性模型,并基于像素的相关性对FCMS算法进行了改进。在设计的算法中,用像素之间的相关性作为邻域像素对中心像素的影响因子,这样可以从很大程度上提高算法的分割效果,消除目标对象的模糊现象。更多还原

【Abstract】 Image segmentation is one of the most important topics in such fields as com-puter vision, image understanding, etc, and is also the core and hot topics in medical imaging. Formally, medical image segmentation is to partition a medical image into non-overlapping different organs or tissues with homogeneous characteristics. By med-ical image segmentation, interesting organ/issue can be retrieved from medical images, and lesions can be analyzed qualitatively and quantitatively. In addition, medical im-age segmentation is the basis of image processing, such as3D reconstruction, and the segmentation results affect the accuracy of reconstruction model directly. Therefore, re-search of medical image segmentation is of important significance and practical value.Due to the complexity of medical images, medical image segmentation has been one bottleneck to restrict medical applications. Considering such phenomena as intensi-ty inhomogeneity, partial volume effect(PVE) and noise, this paper investigates medical image segmentation based on fuzzy clustering algorithms. Compared with other seg-mentation methods, fuzzy clustering is one typical "soft" algorithm. In fuzzy clustering algorithm, pixels can belong to different organs or tissues concurrently with different membership degrees, and can deal with PVE in medical images. When applied in im-age segmentation, fuzzy clustering method can retain as much information as possible from the original image, and thus can retrieve good results. However, since spatial in-formation is not considered, traditional fuzzy clustering algorithm is sensitive to noise and artifacts in medical images, and performs poor in noisy images. Moreover, the ef-ficiency is very low, and cannot satisfies the requirement of medical image processing. For these two problems, the research of this paper are performed from the following aspects:(1) Aiming at low efficiency of FCM, this paper analyzes this problem from the aspect of cluster center computation. In this paper, low efficiency is partly due to the fact that cluster center is computed on the basis of all pixels. Based on this, one stratified segmentation technology is proposed in this paper, in which cluster centers are supposed to be based on pixels in corresponding organ/issues, unrelated with pixels belonging to other clusters. In the proposed algorithm, threshold technique is adopted to partition the medical image into different organ/issue roughly, and then FCM is performed to revise the rough results, which will decease the computation and thus improve the efficiency.(2) This paper investigated the real-time segmentation of medical image segmenta-tion, which can solve low efficiency of FCM and poor result of FCM-related algorithm-s. In this paper, current segmentation algorithms originating from FCM are analyzed deeply, and low efficiency of FCM is due to the fact that the hidden information is not mined sufficiently. Based on this suppose, one improved algorithm based on histogram is proposed, named HisFCM. Peak detection is adopted to partition the histogram of the given image into different intervals, and suppose clustering centers are decided by cor-responding intervals, unrelated to others. Based on these intervals, FCM is performed on the histogram, which can improve the efficiency greatly. Generally, it needs less than0.1s to perform segmentation, which can satisfy the real-time requirement of medical image segmentation.(3) Image segmentation based on neighbor information is proposed. FCMS is one improved algorithm to enhance the segmentation results of FCM, and one constant a is assigned to reflect the impact of neighbor pixels on central one. By such processing, the algorithm can resist the effect of noise. However, when images without noise are segmented, the pixels along the edge maybe misclassified, and fuzzy phenomena appear. This paper proposed the relevance model based on spatial information and intensity information. Then pixel relevance is adopted to replace the impact factor of neighbor pixels on central one. By such, segmentation results can be improved greatly, and fuzzy phenomena can also be eliminated.更多还原