基于超体素的颅内出血区域分割研究

发布时间：2018-01-31 23:19

本文关键词： 图像分割颅内出血图割超体素半监督　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：颅内出血(ICH)是最严重的急性脑血管疾病之一,也是急性神经紊乱疾病,如偏瘫等的重要发病诱因。对于临床治疗来说,颅内出血的分割具有重要意义。因此,如何应用CT图像诊断颅内出血已成为了脑医学领域最热门的研究课题之一。在传统的医学图像分析中,医学从业人员主要依靠手工测绘以及自身的经验,通过估计得到对于颅内出血情况的判断。图像分割技术的引入,大大减轻了医学从业人员的工作负担,分割得到的量化数据也为医学从业人员提供了精确的诊断依据。近年来各种图像分割算法接连提出,其中基于超像素,图论和半监督学习的图像分割算法,由于其良好的分割性能已经成为众多研究人员关注的热点。超像素算法通过依据像素之间特征的相似性将图像划分为小区域,减少了图像中的冗余信息,在很大程度上降低了后续图像处理过程的复杂度。基于图论的图像分割算法则通过将图像特性对应于图论特性,将图像分割问题转化为网络图的分割问题,通过将图像的全局分割与局部信息处理相结合,减少了'图像离散化造成的误差,从而可获得良好的分割结果。在有标签数据稀缺的情况下,基于半监督学习的图像分割方法可以利用大量的无标签数据增强分割结果。本文围绕颅内出血区域分割以及超像素的特性展开研究,重点研究了基于超像素的图割算法和基于超像素的Tri-training算法在颅内出血区域分割领域的应用。本文的主要工作和贡献如下:1、研究了图像分割领域中常用的算法及各自的应用范围,简要介绍了医学图像分割的特点以及颅内出血(ICH)的发病原理以及CT成像特点;2、详细介绍了超像素算法的分类,基本原理以及优缺点,基于现有的超像素算法,结合颅内出血区域分割这一应用场景,提出了一种新的超体素算法;3、详细介绍了图割算法的基本原理以及优缺点,针对基于图割的图像分割算法中人工参与和模型估计不足的问题,提出了一种基于高斯混合模型(GMM)的有监督图割算法。该算法根据医学图像的特性,利用已有的有标签数据为先验知识,通过GMM算法建立前景及背景的模型,使得基于图割的图像分割算法能够实现全自动分割。4、详细介绍了 Tri-training算法的基本原理以及优缺点。结合超体素的特性,针对医学图像分割中有标签样本获取困难这一情况,提出了一种基于超体素和Tri-training算法的颅内出血区域分割算法。该算法利用医学图像处理领域中存在的少量有标签数据和大量无标签数据,实现了医学图像的自动分割。
[Abstract]:Intracranial hemorrhage (ICH) is one of the most serious acute cerebrovascular diseases, and is also an important cause of acute neurological disorders, such as hemiplegia. The segmentation of intracranial hemorrhage is of great significance. Therefore, how to diagnose intracranial hemorrhage with CT image has become one of the hottest research topics in the field of brain medicine. Medical practitioners mainly rely on manual mapping and their own experience, through the estimation of intracranial hemorrhage, the introduction of image segmentation technology, greatly reduce the workload of medical practitioners. In recent years, a variety of image segmentation algorithms have been proposed one after another, which is based on super-pixel, graph theory and semi-supervised learning image segmentation algorithm. Because of its good segmentation performance has become the focus of attention of many researchers. Super pixel algorithm divides the image into small regions according to the similarity of the features between pixels, which reduces the redundant information in the image. Image segmentation algorithm based on graph theory can transform image segmentation problem into network image segmentation problem by mapping image characteristics to graph theory characteristics. By combining global image segmentation with local information processing, the error caused by 'image discretization' can be reduced, and good segmentation results can be obtained. The image segmentation method based on semi-supervised learning can use a lot of unlabeled data to enhance the segmentation results. This paper focuses on the segmentation of intracranial hemorrhage region and the characteristics of super-pixel. This paper focuses on the application of super-pixel based image cutting algorithm and super-pixel based Tri-training algorithm in the field of intracranial hemorrhage region segmentation. The main work and contributions of this paper are as follows: 1. The common algorithms in the field of image segmentation and their application fields are studied. The characteristics of medical image segmentation, the pathogenesis of intracranial hemorrhage (ICH) and the characteristics of CT imaging are briefly introduced. 2. The classification, basic principle, advantages and disadvantages of the super-pixel algorithm are introduced in detail. Based on the existing super-pixel algorithm and the application scene of intracranial hemorrhage region segmentation, a new hypervoxel algorithm is proposed. 3. The basic principle, advantages and disadvantages of graph cutting algorithm are introduced in detail, aiming at the problems of artificial participation and insufficient model estimation in image segmentation algorithm based on graph cutting. In this paper, a supervised graph cutting algorithm based on Gao Si mixed model (GMMM) is proposed. According to the characteristics of medical images, the algorithm uses the existing tagged data as the prior knowledge. The model of foreground and background is established by GMM algorithm, so that the image segmentation algorithm based on graph cutting can realize automatic segmentation. 4. The basic principle, advantages and disadvantages of Tri-training algorithm are introduced in detail. According to the characteristics of hypervoxel, it is difficult to obtain tag samples in medical image segmentation. A region segmentation algorithm for intracranial hemorrhage based on hypervoxel and Tri-training algorithm is proposed, which utilizes a small amount of labeled data and a large amount of untagged data in the field of medical image processing. The automatic segmentation of medical image is realized.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R743.34;TP391.41

【参考文献】