脑认知状态分类识别方法的研究和应用

Research and Application of Brain Cognitive States Classification and Recognition Method

作者: 专业:计算机应用技术 导师:孟军 年度:2010 学位:硕士  院校: 大连理工大学

Keywords

Functional Magnetic Resonance Imaging, Latent semantic Analysis, Principal Components Analysis, K Nearest Neighbor, Hidden Markov Model

        神经信息学是集神经学、信息学以及计算机科学于一体的跨学科研究领域。随着科学的进步和新技术的应用,产生了海量的脑图像数据。近年来,应用高效的数据挖掘算法对这些数据进行分析处理,揭示数据背后隐藏的信息和规律,研究脑认知的工作机理,成为新的科研热点。本文着重讨论在没有医学先验知识情况下,对功能磁共振成像(functional Magnetic Resonance Imaging, fMRI)数据进行预处理、特征降维、特征提取,并在此基础上对隐含脑认知状态进行分类识别的模型。首先介绍了基于统计参数图软件包(Statistical Parametric Mapping, SPM)对fMRI图像进行预处理的详细步骤,并对预处理后的数据进行了初步降维。给出了两种数据驱动的特征抽取方法:潜在语义分析基于奇异值分解,将高维向量投影到低维潜在语义空间中,不仅可以实现有效降维,还能捕获数据的潜在信息;主成分分析以特征值分解为基础,消除多元变量间的相关性,用尽可能少的主成分,甚至仅用第一主成分,就能有效表示原始数据。实验证明,这两种方法可以将fMRI数据降至一维向量,分别用12个元素和8个元素表示。在上述工作的基础上,讨论了两种不同的脑认知状态分类识别模型。在应用潜在语义分析和主成分分析获得特征向量之后,使用改进的K最近邻算法,基于相似度,对待分类样本进行识别分类,取得了84.7%的分类准确率。考虑到fMRI数据的时间序列特性,使用粒子群算法改进了隐马尔科夫模型的训练过程,在应用主成分分析抽取各图像第一主成分之后,基于改进的模型对fMRI图像时间序列进行分类,取得了77.6%的分类准确率。实验证明,这两种机器学习方法,在基于fMRI的脑认知状态分类应用中的取得了很好的效果。
    Neuroinformatics combines neuroscience and computational science and informatics. Massive brain images have been produced as new techniques applied. Applying data mining algorithms to analyze brain images and discover brain cognition mechanism has become an important research area.In this paper, we focus on fMRI data preprocessing, feature reduction and extraction, and modeling to classify hidden brain cognitive states without any domain knowledge. fMRI images obtained in scanning experiment can’t be used directly, which means image registration and standardization should be accomplished first of all. Preprocessing steps based on Statistical Parametric Mapping are discussed in this paper, and preliminary dimension reduction is achieved after that.Two kinds of data-driven feature extraction algorithms are proposed:Latent Semantic Analysis is based on Single Value Decomposition, by which high-dimensional vector can be mapped into low-dimensional latent semantic space. Using LSA, not only an effective dimension reduction can be achieved, but also the potential information of fMRI data can be captured. Principal Component Analysis is based on Eigenvalue Decomposition, by which the correlation among multiple variables can be eliminated. Using PCA, original fMRI data can be effectively represented by several principle components as few as possible, or even just the first principal component. Experimental results show that these two methods can reduce vector dimension and extract vector feature without prior knowledge.Two kinds of brain cognitive states classification model are introduced based on the above work. After feature extraction by LSA and PCA, based on similarity, samples to be treated are classified by an improved K-nearest neighbor algorithm. Considering the time series characteristics of fMRI data, a swarm-based Hidden Markov Model is proposed. After extracting the first principal component of each fMRI image by PCA, the improved HMM is applied in fMRI time series classification. Experimental results indicate that these two methods performance very well in brain cognitive states classification and recognition.
        

脑认知状态分类识别方法的研究和应用

摘要4-5
Abstract5
1 绪论8-11
    1.1 研究背景及意义8
    1.2 国内外研究现状8-9
    1.3 本文的工作9-10
    1.4 论文的组织10-11
2 功能磁共振成像11-22
    2.1 大脑结构功能及其成像技术11-13
        2.1.1 大脑的结构和功能11-12
        2.1.2 脑功能成像技术12-13
    2.2 功能磁共振成像技术13-16
        2.2.1 fMRI的成像原理及其数据特点13-14
        2.2.2 fMRI实验设计方法14-15
        2.2.3 本文采用的实验数据15-16
    2.3 fMRI图像分类识别基本模型16-18
    2.4 fMRI数据预处理18-21
        2.4.1 SPM简介18-19
        2.4.2 SPM预处理19-20
        2.4.3 SPM统计20-21
    2.5 本章小结21-22
3 数据驱动的fMRI特征降维和提取22-38
    3.1 实验数据初步降维22-23
    3.2 fMRI数据分析方法23-25
        3.2.1 模型驱动方法23
        3.2.2 数据驱动方法23-25
    3.3 基于潜在语义分析的特征抽取25-29
        3.3.1 潜在语义分析25-27
        3.3.2 潜在语义分析降维实例27-29
        3.3.3 基于LSA的fMRI特征抽取29
    3.4 基于主成分分析的特征抽取29-34
        3.4.1 主成分分析29-31
        3.4.2 主成分分析降维实例31-33
        3.4.3 基于PCA的fMRI特征抽取33-34
    3.5 实验结果和分析34-37
    3.6 本章小结37-38
4 基于KNN的孤立认知状态分类38-46
    4.1 脑认知状态分类38-39
    4.2 KNN分类模型39-42
        4.2.1 KNN算法思想39-41
        4.2.2 基于类中心的KNN算法41-42
    4.3 实验结果和分析42-45
    4.4 本章小结45-46
5 基于HMM的认知状态时间序列分类46-56
    5.1 HMM分类模型46-48
        5.1.1 HMM模型表示46-48
        5.1.2 HMM模型应用48
    5.2 PSO优化算法48-49
    5.3 基于群优化的HMM训练模型49-53
        5.3.1 迭代终止策略50-51
        5.3.2 粒子编码模式51-52
        5.3.3 参数归一化处理52-53
        5.3.4 适应度评价函数53
    5.4 实验结果和分析53-55
    5.5 本章小结55-56
结论56-57
参考文献57-59
攻读硕士学位期间发表学术论文情况59-60
致谢60-62
        下载全文需10


本文地址:

上一篇:基于虚拟现实技术的心脏建模方法研究
下一篇:微型燃料电池环氧封装结构蠕变特性影响研究

分享到: 分享脑认知状态分类识别方法的研究和应用到腾讯微博           收藏
评论排行
公告