基于描峰聚类的动态脑功能网络状态划分

Dynamic functional brain network division based on density peak clustering

目的近年来脑功能网络的动态属性分析已经成为脑功能研究的热点,脑功能网络状态划分则是脑功能网络动态属性分析的重要方面,目前国际上广泛采用的脑功能网络状态划分策略是k均值聚类算法,而k均值聚类算法存在两个缺陷。而描峰聚类(density peak clustering)算法能直观展现合理的类别数,从而有效解决k均值聚类中k值难以确定的问题。本文拟基于动态功能连接(dynamic functional connectivity,DFC)的脑功能网络状态划分,为脑功能网络划分探索新的模型。方法基于61位成年人静息态功能磁共振成像(resting state functional magnetic resonance image,rs-f MRI)数据,采用滑动窗口计算方法构建功能连接矩阵。基于多种距离度量使用多维尺度分析算法对其进行有效降维,通过描峰聚类算法进行脑功能网络状态划分,使用脑功能网络划分常用的状态模式图和聚类决策图进行结果的校验。结果基于余弦距离、相关系数以及Spearman等描述相似性的距离度量进行降维,得到的结果生理意义较为明确,且有效功能网络状态数为3~5。另外,脑区之间松散联系的网络状态比其他网络状态更频繁地发生。结论描峰聚类算法足以对个体脑功能连接随时间的动态波动进行状态划分,这可为脑功能网络划分研究提供新的思路。

Objective In recent years,the dynamic attribute analysis of functional brain network has become a hot topic in functional brain research.The network state division is an important aspect of network dynamic attribute analysis.At present,the internationally widely used network state division strategy is k-means clustering algorithm which has two defects.Density peak clustering can visually display the reasonable number of categories,which can effectively solve the problem that k-values in k-means clustering are difficult to determine.We intend to explore a new model for network state partitioning based on dynamic functional connectivity(DFC).Methods Based on resting state functional magnetic resonance image(rs-f MRI)data of 61 adults,a sliding window calculation method is used to construct a functional connectivity matrix.Multidimensional scale analysis algorithm is used to effectively reduce dimension based on multiple distance metrics,and network state division is performed by density peak clustering.State pattern graph and cluster decision graph are used to verify the results.Results Based on cosine distance,correlation coefficient,Spearman and other distance measures to describe the similarity,the physiological significance of the results are clear,and the number of effective functional network states is 3 to 5.In addition,network states that are loosely connected between brain regions occur more frequently than other network states.Conclusions The state of functional brain network dynamic fluctuation with time can be divided by density peak clustering algorithm,which can provide new ideas for functional brain network partition.