脑图像多层次智能分类算法研究及其应用

发布时间：2018-01-09 01:10

本文关键词：脑图像多层次智能分类算法研究及其应用　出处：《中国科学院长春光学精密机械与物理研究所》2017年博士论文　论文类型：学位论文

【摘要】：随着医学影像技术的不断成熟和发展,基于脑图像的大脑结构和脑功能分析,已经成为近年来的研究热点。采用传统的手动勾画方式进行特征提取和脑图像分析,极大的增加了临床医生的负担。通过机器学习的方法,自动提取脑图像中的特征,用于脑疾病的诊断和预测,同时,找出与疾病相关的影像学标记,已成为研究趋势。自动的特征提取方法与智能的分类算法,可以提高医生对脑疾病诊断的效率,具有较高的应用价值。在实现对脑疾病辅助诊断的过程中,首要的任务是提取出能使分类器性能最优的特征。研究脑图像自动特征提取方法,特别是建立出能够全面反映大脑结构信息的高维度、多层次特征,是分类算法研究中的难点与热点问题。目前,脑图像研究中使用的特征大多数是基于体素的或是基于表面的,这些特征大多是单层次的,不能够全面、综合的表达脑疾病相关的大脑结构信息。本论文提出一种脑图像多层次智能分类算法,用于脑疾病的辅助分析。本论文进行的主要研究工作和创新点如下:(1)通过对自动特征提取方法进行研究,提出一种基于脑图像的多层次智能分类算法,有助于找出大脑结构的影像学标记。多层次特征包括低层次感兴趣区域(Region of interest,ROI)特征(大脑体积和皮层厚度等)和高层次脑网络特征(ROI之间的功能连接)。采用基于Filter和基于Wrapper的混合特征选择方法分别对两类特征进行降维。针对每一种特征,分别使用径向基核函数(Radial basis function,RBF)构造核矩阵,再通过适当的权重因子将两类核矩阵集成为一个基于支持向量机(Support vector machine,SVM)的多核分类器。采用嵌套的交叉验证方法进行算法的评估。多层次特征能够更全面的表达大脑结构信息,分类算法具有较强的适用性,可以应用于不同脑疾病的分析和分类。(2)应用自动的图像处理方法,进行了2型糖尿病脑结构影像学标记物的分析,找出了具有统计学意义的影像学标记,可以提升对2型糖尿病的早期辨识。2型糖尿病是一种常见的代谢性疾病,会对脑组织造成不可逆的损伤,引起认知障碍等并发症。通过影像学方法对患者大脑结构进行检测,有助于对2型糖尿病的诊断和治疗。之前的影像学研究大多数使用的是单一的大脑体积或者皮层厚度等测量值,不能够综合的反映2型糖尿病大脑结构的改变。本论文通过使用性能较好的脑组织分割算法和脑皮层重建算法,可以准确计算出灰质体积、皮层厚度和皮层表面积等测量值。这三种测量值可以从不同层面反映出大脑的结构信息,如,灰质体积表示神经元的总体数量,皮层厚度表示纵向排列的神经元数量,皮层表面积表示横向排列的神经元数量。同时研究这三种测量值对于全面分析2型糖尿病大脑结构改变有重要意义。(3)应用多层次智能分类算法进行了自尊程度与大脑结构的关联分析。为了更好的理解自尊程度这一复杂的认知心理学过程,需要对大脑自我认知的神经机制进行充分的研究。目前基于自尊程度的神经影像学研究,大都采用的是基于全脑形态学分析方法或基于体素的脑体积分析方法,只能对从先验信息中得到的特定脑区进行分析。虽然这些方法初步揭示了自尊程度与大脑结构之间的关系,但是不能很好的解释神经网络活动与自尊程度的关联性。本论文采用多层次智能分类算法,具有较高的分类准确率(96.66%)、特异性(99.77%)和敏感性(95.67%),同时,找出了对自尊程度敏感的脑区,不仅能够弥补目前研究方法的不足,而且能够同时提供大脑形态学信息和roi之间的功能连接信息,可以帮助研究人员更好的理解不同自尊程度的大脑模型,对临床研究和科学研究有重要意义。(4)虽然目前国内外有一些常用的医学图像处理软件平台可以完成图像的自动处理、进行脑图像的分析以及辅助脑疾病的诊断,但是还不能很好的从图像中自动提取出所需要的多层次特征。本论文中开发了brainlab全自动脑图像处理与分析系统,能够对脑图像进行自动的处理和多层次特征提取,辅助医生进行脑疾病的早期诊断和分析,在临床应用中具有重要意义。
[Abstract]:With the development of medical imaging technology, analysis of brain structure and brain function of brain based on image, has become a research hotspot in recent years. The traditional way of manual delineation of feature extraction and analysis of brain image, greatly increased the burden of clinicians through a machine learning method. The automatic feature extraction of brain images for the prediction and diagnosis of brain diseases, and find out the disease associated imaging marker, has become a research trend. The classification algorithm and intelligent automatic feature extraction method, can improve the medical diagnosis of brain disease efficiency, and has high application value. In the process of implementation of the diagnosis of brain diseases first, the task is to extract the feature classifier optimal performance. Research on automatic feature extraction method of brain images, especially the establishment of a fully reflect the high dimensional brain structure information Of multi level feature, is a hot issue and difficulty in classification algorithm research. At present, the use of the characteristics of brain images are based on a voxel or is based on the surface, these features are mostly single level, can not be a comprehensive expression, brain disease related brain structures. This paper presents comprehensive information a brain image multilevel intelligent classification algorithm for the analysis of auxiliary brain diseases. The main research work and innovations are as follows: (1) based on the research of automatic feature extraction method, puts forward a multi-level intelligent classification algorithm based on brain image, helps to find out the structure of the brain imaging marker. Multi level features include the low level of region of interest (Region of, interest, ROI) features (brain volume and cortical thickness etc.) and the high level of brain network characteristics (ROI between functional connectivity). Based on Filter and based on Wrap A hybrid feature selection method of per were used to reduce the dimensionality of the two kinds of features. For each feature, respectively using RBF kernel function (Radial basis, function, RBF) to construct the kernel matrix, then the appropriate weight factor will be two types of nuclear matrix into one based on support vector machine (Support vector machine, SVM) the multi kernel classifier evaluation. Using cross validation method for nested algorithm. Multi level feature can brain structure information expression more comprehensive, classification algorithm has strong applicability, can be applied to the analysis and classification of different brain diseases. (2) application of the automatic image processing method, analyzed the structure of the brain type 2 diabetes imaging markers, were found statistically significant imaging markers for early identification, can improve the type.2 diabetes type 2 diabetes is a common metabolic disease, will not cause of brain tissue Reversible injury, complications of cognitive disorders. To detect the brain structure of the patients through the imaging method, is helpful to the diagnosis and treatment of type 2 diabetes. Previous imagingstudies most use single brain volume or cortical thickness measurements, cannot reflect the change of brain structure in type 2 diabetes mellitus. Through the use of better performance of brain tissue segmentation algorithm and cortical reconstruction algorithm can accurately calculate the gray matter volume, the thickness of cortex and cortex surface area measurements. The three kinds of measurements can reflect the structural information from different aspects such as the brain, said the total number of neurons in gray matter volume, cortical thickness the number of vertical arrangement of neurons, cortical surface area indicates the number of horizontal array of neurons. At the same time of the three kinds of measurements for a comprehensive analysis of type 2 diabetes brain. The change has important significance. (3) the application of multilevel intelligent classification algorithm are analyzed and correlated self-esteem and brain structure. In order to better understand the self-esteem of the cognitive psychology process, need the neural mechanism on the brain self cognition are fully studied. Based on the current neuroimaging study of self-esteem, mostly the full brain morphological analysis method based on the method of analysis of brain volume or voxel based analysis of specific regions of the brain can only be obtained from prior information. Although these methods reveal the relation between self-esteem and the degree of brain structure, but the correlation explain the activity of neural network and the level of self-esteem is not very good. This paper adopts multilevel intelligent classification algorithm has high classification accuracy (96.66%), specificity (99.77%) and sensitivity (95.67%), at the same time, to find out the level of self-esteem Sensitive areas of the brain, not only can make up for the current research methods, but also can provide information between brain morphology and the function of ROI connection information, the brain model understanding of different levels of self-esteem can help researchers better, has important significance for clinical research and scientific research. (4) although there are some medical images processing software platform can complete the automatic image processing commonly used, diagnosis of brain image analysis and auxiliary brain diseases, but also not very good from the image to extract multi-level features needed. This paper developed a BrainLAB automatic brain image processing and analysis system to brain image processing and multi-level automatic feature extraction, early diagnosis and analysis of doctor assisted brain diseases, has important significance in clinical application.

【学位授予单位】：中国科学院长春光学精密机械与物理研究所
【学位级别】：博士
【学位授予年份】：2017
【分类号】：R741.044;TP391.41

【相似文献】