基于模糊聚类的医学图像分割算法研究及设计

发布时间：2018-04-13 06:09

本文选题：模糊聚类 + 医学图像分割　；参考：《西南科技大学》2017年硕士论文

【摘要】：图像分割是图像处理领域的重要研究内容,普遍应用于医学、气象、计算机视觉、军事、遥感等诸多研究领域。医学图像分割就是对医学图像进行有意义的划分,使其成为指定个数的相异区域,每个区域内都保持一致性,区域之间尽可能地不同,并且各区域互不相交。医学图像分割为特征提取与识别、三维可视化、病理分析与诊断等提供有用的信息。医学图像分割的主要目的是对医学图像进行有医疗价值的划分,并提取出特定的区域,便于医生制定医疗方案,开展疗效评估等。医学图像在成像时会受到分辨率、光照条件等影响,因此存在不确定性,而模糊处理技术正好适用于此类问题。本文深入研究了模糊聚类基本理论以及其它人工智能技术,分析了前人算法中存在的问题以及在医学图像分割中遇到的困难,提出了几种改进的模糊聚类算法,并且应用于医学图像分割中。本文主要获得的创新性成果如下:(1)在现有的相关评价标准的基础上,本文提出了新的聚类算法评价标准:聚类中心变化值指数dv。作为衡量一个聚类算法优劣的新标准,dv指数适用于所有基于聚类中心初始化的聚类算法,不仅仅是模糊聚类,也不仅仅是图像分割;关注的是以何种方法寻找的初始聚类中心更好,从结果的角度或者算法整体的角度评判一个聚类算法的优劣。(2)在快速模糊聚类算法FFCM和En FCM基础上,本文提出了一种更加高效的医学图像分割算法。该算法先将图像的直方图与高斯模板卷积,并进行峰值检测,得到c个峰值的横坐标,以此为基础进行区间划分,在各区间范围内初始化并且更新聚类中心,然后对原图像进行均值滤波,最后根据像素值所属区间完成分割。实验证明,在保证分割质量的前提下,可将运行时间减少4%以上,优于目前的快速分割算法。(3)充分挖掘利用图像直方图的区间信息,得出了两种新的基于模糊聚类的医学图像分割算法。首先用区域分裂求初始聚类中心或者区域迭代均值化求初始聚类中心,简化隶属度的初始化方式,采用En FCM的方法更新隶属度,并提出了对隶属度进行选择性计算的策略,进一步减少了算法的计算量,提高了运行效率;再由隶属度的变化更新相应区间的范围,限制聚类中心的计算于每个区间中,直到前后两次聚类中心的变化值小于指定的阈值,迭代结束,完成分割。通过实验证明,两种新算法均不同程度地优于前述相关算法,尤其是第二种新算法在其他参数表现稳定的情况下,效率也是最优的。(4)利用隶属度的选择性计算来改进qjjc FCM算法和gsav FCM算法。前者改进后优于改进前,而后者改进后的表现不佳。我们分析了原因,或与其寻找初始聚类中心的方式和区间划分方法有关。
[Abstract]:Image segmentation is an important research content in the field of image processing. It is widely used in many fields such as medicine, meteorology, computer vision, military affairs, remote sensing and so on.Medical image segmentation is a meaningful division of medical image, making it a designated number of different regions, each region is consistent, the regions are as different as possible, and the regions do not intersect each other.Medical image segmentation provides useful information for feature extraction and recognition, 3D visualization, pathological analysis and diagnosis.The main purpose of medical image segmentation is to divide the medical image with medical value, and extract the specific region, which is convenient for doctors to formulate medical plan and evaluate the curative effect.Medical images are affected by resolution, illumination conditions and so on, so there is uncertainty, and fuzzy processing technology is suitable for this kind of problems.In this paper, the basic theory of fuzzy clustering and other artificial intelligence techniques are deeply studied, the problems existing in previous algorithms and the difficulties encountered in medical image segmentation are analyzed, and several improved fuzzy clustering algorithms are proposed.And it is applied to medical image segmentation.The main innovative results obtained in this paper are as follows: (1) on the basis of the existing evaluation criteria, a new evaluation criterion of clustering algorithm is proposed: the index of variation value of clustering center (dv).As a new standard for evaluating the advantages and disadvantages of a clustering algorithm, the DV index is suitable for all clustering algorithms based on clustering center initialization, not only fuzzy clustering, but also not only image segmentation;The focus is on which method is better to find the initial clustering center. Judging the advantages and disadvantages of a clustering algorithm from the point of view of the result or the whole algorithm) based on the fast fuzzy clustering algorithms FFCM and en FCM.In this paper, a more efficient medical image segmentation algorithm is proposed.The algorithm firstly convolution the histogram of the image with Gao Si template, and detect the peak value, get the transverse coordinate of the c peak value, and then divide the interval, initialize and update the cluster center in the range of each interval.Then the original image is filtered by the mean value, and the segmentation is completed according to the interval of the pixel value.Experiments show that the running time can be reduced by more than 4% under the premise of ensuring segmentation quality, which is better than the current fast segmentation algorithm.Two new medical image segmentation algorithms based on fuzzy clustering are proposed.At first, the initial cluster center is obtained by region splitting or region iteration mean value, the initialization of membership degree is simplified, the membership degree is updated by en FCM method, and the strategy of selective calculation of membership degree is put forward.It further reduces the computational cost of the algorithm and improves the running efficiency. Then the range of the corresponding interval is updated by the change of the membership degree, and the calculation of the cluster center is limited to each interval, until the change value of the two clustering centers is less than the specified threshold.At the end of the iteration, the partition is completed.The experiments show that the two new algorithms are better than the previous algorithms to some extent, especially the second new algorithm is stable in the case of other parameters.Qjjc FCM algorithm and gsav FCM algorithm are improved by the selective calculation of membership degree.The former is better than the former, while the latter is not good after the improvement.The reason is analyzed, or it is related to the method of searching for the initial cluster center and the method of interval partition.
【学位授予单位】：西南科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】