基于fMRI数据的脑功能网络聚类研究

发布时间：2018-03-03 02:31

本文选题：功能磁共振成像　切入点：脑功能网络　出处：《南京理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：人的大脑由许多的神经元组成,神经元之间互相作用,以完成人脑的各种功能。因为人脑中的神经元数量非常庞大,单独研究起来非常不方便。目前,应用最广泛的方法是基于血氧水平依赖(BOLD)的功能磁共振成像(fMRI)信号,该信号不仅可以考察受试者在静息态下的功能性脑网络,也可以研究基于任务的动态功能性脑网络。传统的分析方法是基于模型的统计方法,但fMRI涉及到脑功能的复杂模式,所以不适用于此类方法,而基于数据驱动的聚类分析则能更科学和客观地分析fMRI数据。本文主要对聚类方法进行研究,使聚类算法能更好的应用于fMRI数据的分析,进而促进脑功能连接的认识。首先,本文通过对使用广泛的K-means算法进行了研究,提出一种基于 DBI的层次初始的K-means 算法(DBI based Hierarchical Initialization K-means,DHIKM),可以选取较好的初始聚类中心和自动地确定聚类个数,并得到了较好的实验结果。由于稀疏表示和结构稀疏表示可以利用信号的稀疏性特征很好地表示fMRI信号,近些年来,许多研究者采用稀疏表示来对fMRI进行研究,并取得了良好的效果。因此,本文在稀疏表示的基础上融合提出的DHIKM算法,对脑功能网络进行聚类研究。通过分析发现,脑图像具有高维数据结构特征,而子空间学习则可以有效降低数据的维度,可以有效地表示fMRI信号,所以采用稀疏子空间学习方法对脑功能网络进行聚类分析。同时,基于空间距离上较近的神经元彼此之间相连的概率较高这一特性,本文提出了邻近自适应的局部尺度稀疏子空间聚类算法(neighboring adaptive local scale based sparse subspace clustering),并得到了较好的成果。
[Abstract]:The human brain is made up of many neurons that interact with each other to perform the functions of the human brain. Because the number of neurons in the human brain is so large, it is very inconvenient to study it alone. The most widely used method is the functional magnetic resonance imaging (fMRI) signal based on Bod, which can not only examine the functional brain network of subjects in resting state. Traditional analysis methods are model-based statistical methods, but fMRI involves complex patterns of brain function, so it is not suitable for this kind of methods. But the data-driven clustering analysis can analyze the fMRI data more scientifically and objectively. In this paper, the clustering method is mainly studied, so that the clustering algorithm can be better applied to the analysis of fMRI data, and then promote the understanding of brain functional connection. In this paper, the widely used K-means algorithm is studied, and an initial K-means algorithm based on DBI is proposed, which can select a better initial clustering center and determine the number of clusters automatically. Because sparse representation and structural sparse representation can be used to represent fMRI signal well, in recent years, many researchers use sparse representation to study fMRI. Therefore, this paper fuses the proposed DHIKM algorithm on the basis of sparse representation, and studies the clustering of brain functional networks. It is found that brain images have high dimensional data structure characteristics. Subspace learning can effectively reduce the dimension of data and effectively represent the fMRI signal, so the sparse subspace learning method is used to cluster the brain functional network. At the same time, Based on the high probability of close neurons being connected to each other in spatial distance, this paper presents a neighborhood adaptive local scale sparse subspace clustering algorithm called nebuling adaptive local scale based sparse subspace clustering, and good results are obtained.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP311.13;R445.2

【相似文献】