【作者】 丁伟东；

【导师】 袁景淇；

【作者基本信息】 上海交通大学 ， 控制理论与控制工程， 2008， 博士

【摘要】 大脑通过神经元动作电位即锋电位进行信号的传递、交流和处理,对神经锋电位活动的记录检测是神经科学研究的前提。神经锋电位信号主要是通过细胞外电极进行记录。然而,单个电极上记录到的信号往往是几个相邻神经元的锋电位与大量噪声的叠加。为了从电极记录信号中获得有用的神经元放电信息,对神经元放电序列进行甄别就尤为重要,有必要把每个神经元发出的锋电位从记录信号中分离出来。模式识别通过数据的先验知识和统计信息来对数据进行分类。本文对模糊聚类和支持向量机等模式识别方法进行深入研究,提出一些简单有效的分类方法解决神经锋电位信号识别中的难点问题。本文的创新性工作有:1.实际检测到的锋电位信号往往包含大量噪声和野值点,针对该问题,提出鲁棒模糊聚类方法提高锋电位的分类精度。在聚类过程中,样本点相对于聚类的模糊隶属度不仅与聚类中心的距离有关,还与样本点局部密度值有关。通过减少具有较小密度值的噪声和野值点的模糊隶属度来降低它们在聚类过程中的影响,同时减少聚类间边界点对于各个聚类的隶属度,使聚类更好地分离开,实现对锋电位信号的准确分类。并在此基础上,提出了改进的模糊聚类有效性评价指标,实现在噪声情况下对锋电位数据聚类个数的识别。2.神经元爆发式放电和记录电极漂移导致单一锋电位聚类形状发生无规则变化,划分聚类方法难以获得满意分类结果。提出基于模糊C均值的层次聚类方法解决这一难题。首先,使用模糊聚类方法将锋电位数据划分为多个关系紧密的小类。然后,通过模糊隶属度来计算次聚类间的连接强度,从整体上衡量划分间的相似性,对小类逐步合并,正确识别复杂的聚类结构。最后,通过样本点与多个中心的平均加权距离来度量类内和类间距离,正确反映出复杂聚类的松紧程度,实现对形状发生无规则变化的锋电位聚类个数地判断。3.针对锋电位信号的叠加问题,提出基于监督分类的模板匹配方法。根据多类支持向量机的分类特点,设计一种新的训练策略,即在标号为负的训练样本中引入人工合成的叠加锋电位波形,对叠加锋电位信号准确判断,然后通过模板提取对检测到的叠加信号进行分离。对于证据理论神经网络分类器,根据其能识别模糊输入向量的特点,通过设置恰当的阈值实现对叠加信号的识别,在分类过程中对这些叠加信号逐步分离。4.对支持向量机及其在锋电位分类中的应用进行深入研究。提出一种基于密度信息的加权支持向量机,根据样本点在高维特征空间密度大小调整其与分类平面的距离,减少噪声和野值点的影响,突出具有较大密度值的重要样本点的作用,从而把训练样本的密度分布信息用于构造最优分类平面。然后,提出典型性支持向量机,根据训练样本的分布特点,在高维特征空间选择一些能代表每一类样本分布信息的样本点,在构造最优分类平面过程中,保证训练样本被正确区分的前提下,最大化与典型性样本的距离,把更多恰当的训练样本分布信息用来构造分类平面,提高支持向量机泛化性能。在对仿真神经锋电位数据的实验中,典型性支持向量机取得了较好的分类结果。更多还原

【Abstract】 The extracellular recording of neural spike activity is a prerequisite for studying information transmission and processing in the brain. The spike recordings are usually obtained with electrodes. However, the recording in single electrode contains spikes from several neurons adjacent to the electrode and a high amount of background noise. Therefore, it is necessary to identify the neural spikes and find out the number of neurons contributing to the electrode recording before further analysis is carried out.Pattern recognition aims to classify data based on either a priori knowledge or on statistical information extracted from the patterns. In this dissertation, the pattern recognition methods of fuzzy clustering and support vector machine (SVM) are studied deeply, several effective classification methods are proposed to deal with the difficulties in spike sorting.The main contributions of this dissertation are summarized as follows:1. A robust fuzzy clustering method is proposed to reduce the influence of noise and outliers in spike sorting. The fuzzy membership degrees are adjusted according to the density values of data points. The noise and outliers with lower density values will have small influence in clustering process. Moreover, the border data between different clusters with low density values will have low memberships to any cluster, which make clusters well separated. In order to obtain the optimal number of clusters with ellipsoidal shapes, an extended fuzzy cluster-validity index is proposed. The fuzzy separation and compactness of clusters are evaluated using the weighted Mahalanobis distances between clusters in the index. The robust fuzzy clustering method is able to classify the real neural spikes with noisy data and outliers.2. In the presence of spike bursts or electrode drift, the spike waveforms generated by single neuron are varying with time, and then the spike clusters will be smeared and have non-convex shapes. An unsupervised hierarchical clustering based on fuzzy C means is proposed to resolve the problem. The initial clusters are obtained by fuzzy clustering method. Considering the interrelations among initial clusters and the complicated structures of spike clusters, the similarity small clusters are merged based on fuzzy membership degrees. The optimal cluster number is obtained by improved Dunn’s index. The index adopts the weighted distances to calculate within-cluster scatter and between-cluster separation, which is suited for clusters with complicated structures.3. A template matching based on supervised classification method is proposed to decompose the overlapping spike waveforms. A new training method is designed for multi-class SVM, the overlapping spikes are identified by introducing synthesized overlapping waveforms into training sets. The detected overlapping spike waveforms are decomposed by template extraction. For evidence-theoretic neural network, the overlapping spikes are directly detected by predetermined thresholds, and then they are decomposed in classification process step by step.4. The SVM and its application in spike sorting are studied deeply. A weighted SVM is proposed based on density information, the density values of training data in feature space is used to adjust the distance from the hyperplane to them. The data with high density will be important to construct the separating hyperplane, while the noisy data with low density will have small influence. And then, the representative SVM is proposed to further improve the classification performance. Some important data are selected according to the distribution characteristic of training data, which should represent the distribution information. The separating hyperplane is constructed by maximizing the distance between the representative vectors and the hyperplane with all the training data being classified correctly. In this way, more useful information of the training data far away from the hyperplane is introduce to reduce the influence of outliers and improve the generalization ability of the learning machine. The proposed methods are applied to the simulated neural spikes and the representative SVM exhibits better classification performance.更多还原