基于多尺度脑网络局部特征的抑郁症分类研究

发布时间：2018-06-18 00:13

本文选题：抑郁症 + 功能磁共振　；参考：《太原理工大学》2017年硕士论文

【摘要】：抑郁症是一种伴有持续的悲伤情感和其他精神病变的一种常见精神疾病,它的发病率高,却难以治愈,是一种潜在的死亡因素,影响着一个个家庭。尽管近年来对于抑郁症的研究从未停歇,它的诊断却还是没有明朗化。神经影像学是诊断精神疾病的一种临床手段。通过静息态核磁共振对人脑的扫描,得到大脑的医学影像,进一步分析病变部位。在此基础上,大脑可以划分为多个功能相对独立的区域,结合复杂网络的知识以这些区域作为节点构建得到一个尺度的脑网络,提取脑网络中的数据进行分类可以作为医学影像辅助诊断精神疾病的一种工具。随着脑网络技术研究的发展,有很多学者将其应用在抑郁症的计算机辅助诊断中。目前,基于抑郁症的脑网络数据进行的大量分类研究多基于单一空间尺度下的脑网络,且使用的特征多为临床指标或脑网络的基础构建元素。一些研究将重点放在特征选择方法的比较和特征的选取上,以期得出辅助诊断抑郁症的最佳方案,对空间尺度影响的研究还远远不够。本文根据先前的研究,在前人的基础上进一步针对脑网络的空间尺度对抑郁症的分类做深入探讨。本文的主要工作如下:第一,不同尺度的脑网络分类比较。采集抑郁症患者和正常被试的静息态磁共振脑影像数据,划分脑区后进一步得到五个不同节点个数的尺度下的脑网络,尺度分别为90,256,497,1003,1501。对不同尺度的脑网络提取局部属性,使用双样本T检验对两类被试的数据做统计分析,选取具有统计显著性的属性作为分类特征,使用支持向量机(Support Vector Machine,SVM)分类后进行比较。第二,分析初始分类结果的影响因素。首先,为了判断有效特征的影响,针对不同尺度的脑网络,将大尺度脑网络下的有效特征逐个替换为小尺度的有效特征,与大尺度脑网络下的有效特征结合并进行分类比较。其次,为了判断特征数目的影响,分别从小尺度脑网络的有效特征中抽取与大尺度有效特征数相同的特征进行分类比较。第三,尺度是否越小越好?为判断不同脑网络尺度对于分类结果的影响,使用最大相关最小冗余(minimal Redundancy Maximal Relevance,mRMR)方法对特征与不同尺度脑网络类标签的相关性及不同尺度脑网络下特征间的冗余度进行分析,并判断不同尺度脑区间的距离及脑区体积对冗余度的影响。第四,结合时间复杂度及以上分析,得出某一尺度脑网络下的相对最优分类器,并得出抑郁症患者的异常脑区。综合考量,分类效果是随着尺度的减小而提升的,然而,经本文分析,尺度并非越小越好,使用1003尺度的脑网络构建分类器对抑郁症的分类效果较佳。
[Abstract]:Depression is a common mental disease with persistent sadness and other mental disorders. Its incidence is high, but it is difficult to cure. It is a potential death factor and affects families. Although research on depression has never stopped in recent years, its diagnosis remains unclear. Neuroimaging is a clinical method for the diagnosis of mental diseases. The medical images of brain were obtained by scanning the brain by resting MRI, and the lesion location was further analyzed. On the basis of this, the brain can be divided into several regions with relatively independent functions, which can be combined with the knowledge of complex networks to construct a scale brain network using these regions as nodes. Extracting data from brain network for classification can be used as a tool of medical image aided diagnosis of mental illness. With the development of brain network technology, many researchers have applied it to computer aided diagnosis of depression. At present, a large number of classification studies based on depression brain network data are based on a single spatial scale of the brain network, and the characteristics of the use of clinical indicators or the basic elements of the brain network. Some studies focus on the comparison of feature selection methods and feature selection in order to obtain the best scheme for the diagnosis of depression. On the basis of previous studies, this paper makes a further study on the classification of depression based on the spatial scale of brain network. The main work of this paper is as follows: first, the classification of different scales of brain network comparison. The resting magnetic resonance imaging data of depression patients and normal subjects were collected and the brain networks with five different nodal numbers were obtained after dividing the brain regions. The local attributes were extracted from different scales of brain network, and the data of two kinds of subjects were statistically analyzed by double sample T test. The attributes with statistical significance were selected as classification features, and the support vector machine support Vector Machine (SVM) was used to classify and compare. Secondly, the factors influencing the initial classification results are analyzed. Firstly, in order to judge the influence of effective features, the effective features of large-scale brain networks are replaced with those of small scale ones one by one, and the effective features of large-scale brain networks are combined with the effective features of large-scale brain networks to classify and compare. Secondly, in order to judge the influence of the number of features, the features which are the same as the large scale effective features are extracted from the effective features of the small scale brain network for classification and comparison. Third, is the smaller the better? In order to judge the influence of different scales of brain network on the classification results, the correlation between features and the labels of different scales of brain networks and the redundancy of features under different scales of brain networks were analyzed by using the method of maximum redundancy minimal redundancy MRs. The influence of the distance between different scales and the volume of brain area on redundancy was evaluated. Fourthly, combining the time complexity and the above analysis, the relative optimal classifier under a certain scale brain network is obtained, and the abnormal brain regions of depression patients are obtained. Overall, the classification effect is improved with the decrease of scale. However, the smaller the scale is, the better the scale is. The classifier based on the brain network with scale 1003 is better for the classification of depression.
【学位授予单位】：太原理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R749.4;TP391.41

【参考文献】