基于矢量量化的立体图像分割及在MRI中的应用

发布时间：2018-09-15 05:11

【摘要】：图像分割是图像处理中的一项关键技术,近几年图像分割技术逐渐向着立体图像处理方向发展,针对医学立体图像的分割研究成为一个重要的研究方向。目前对立体图像分割的研究中,在邻域结构信息的充分利用、自动确定分割数目以及空域特征提取与分割等方面尚需进一步完善。对医学立体图像分割的研究仍是一项具有挑战性的课题,具有重要的研究意义和临床实际应用价值。本文针对医学MRI立体图像的分割进行了深入的研究,主要研究工作如下：本文对利用矢量量化技术实现图像分割的方法进行了研究,利用矢量量化过程对图像进行以局部区域(图像子块)为单元的分割,提出了图像分割的矢量量化方法。该分割方法不仅针对图像像素灰度信息进行,还利用了像素邻域结构信息,符合人类视觉对外界信息的认知过程。在实现图像子块矢量量化的过程中,采用基于SOM神经网络的方法完成码本设计,并通过利用基于最小化类内离散度与类间离散度比值的方法求取最佳码本尺寸,据此自适应确定图像的分割数目。针对医学MRI立体图像的分割,提出了一套基于矢量量化的MRI立体图像分割方法。该方法以立体图像的空间子块为基本单元,将矢量量化应用于三维数据的分割之中。针对MRI立体图像的特点,设计了立体图像的层间插值算法、空间子块的边缘模式检测算法；在利用SOM神经网络获取量化码本过程中,设计了分层分割和整体分割两种方式；利用矢量量化过程完成对由空间子块构造的矢量进行的自适应立体分割。实际中,将提出的立体图像分割方法应用于人脑MRI立体图像的分割之中,分别以IBSR图像库和BrainWeb图像库中的仿真立体图像和真实立体图像为样本进行实验和结果分析。在立体图像值域分割基础上,本文通过对其空域特征进行分析,提出了一种立体图像的空域分割方法。该方法首先通过对立体图像空间连通性的检测得出值域分割结果中各个空间体在三维空间中的连通关系,并据此将值域分割结果进行进一步分割；然后通过提取各空间体相应的空域几何参数,对得到分割结果进行定量描述,完成对MRI立体图像的最终分割。通过在人脑MRI立体图像值域分割基础上进行的空域分割实验,验证了提出的空域分割方法的有效性,并将得出的人脑各部分的定量信息和病灶随时间变化的空域几何参数,应用于实际临床医学研究和治疗中。
[Abstract]:Image segmentation is a key technology in image processing. In recent years, image segmentation technology has gradually developed towards stereo image processing, and the research on medical stereo image segmentation has become an important research direction. At present, in the research of stereo image segmentation, it is necessary to make full use of neighborhood structure information, determine the number of segmentation automatically, and extract and segment spatial features. The research on medical stereo image segmentation is still a challenging subject with important research significance and clinical application value. In this paper, the segmentation of medical MRI stereo image is deeply studied. The main research work is as follows: in this paper, the method of image segmentation based on vector quantization is studied. Using the vector quantization process to segment the image with local region (image sub-block) as the unit, a vector quantization method for image segmentation is proposed. The segmentation method not only aims at the gray level information of image pixels, but also utilizes the pixel neighborhood structure information, which accords with the cognition process of human vision to the outside information. In the process of realizing image subblock vector quantization, the codebook design is accomplished based on SOM neural network, and the optimal codebook size is obtained by using the method based on minimizing the ratio of intra-class dispersion to inter-class dispersion. According to this, the number of image segmentation is determined adaptively. A set of MRI stereo image segmentation method based on vector quantization is proposed for medical MRI stereo image segmentation. In this method, the spatial subblock of stereo image is taken as the basic unit, and vector quantization is applied to the segmentation of 3D data. According to the characteristics of MRI stereo image, the interlayer interpolation algorithm of stereo image and the edge pattern detection algorithm of space sub-block are designed, and two methods of hierarchical segmentation and global segmentation are designed in the process of obtaining quantization codebook by using SOM neural network. The vector quantization process is used to realize the adaptive stereo segmentation of the vector constructed by the space subblock. In practice, the proposed stereo image segmentation method is applied to the segmentation of human brain MRI stereo image. The simulation stereo image and the real stereo image in the IBSR image library and the BrainWeb image database are taken as the samples and the results are analyzed respectively. On the basis of stereo image range segmentation, a spatial segmentation method of stereo image is proposed by analyzing its spatial features. Firstly, by detecting the spatial connectivity of the stereo image, the connectedness of each spatial body in the three-dimensional space is obtained, and then the range segmentation results are further segmented. Then, by extracting the corresponding spatial geometric parameters of each spatial volume, the segmentation results are quantitatively described, and the final segmentation of the MRI stereo image is completed. Based on the range segmentation of human brain MRI stereo image, the validity of the proposed spatial segmentation method is verified, and the quantitative information of each part of the human brain and the spatial geometric parameters of the lesion changing with time are obtained. It is applied to clinical medical research and treatment.
【学位授予单位】：大连理工大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】