基于相似度矩阵约减的仿射聚类fMRI数据分析

摘要

Affinity Propagation Clustering(APC)method shows its limitations in time complexity,data storage and clustering results while handling massive functional Magnetic Resonance Imaging(fMRI)data.Aiming at these problems,this paper proposes a new method named SDAPC,which combines Sparse APC(SAPC)with similarity matrix reduction.It starts from sparse approximation on fMRI data,continues with the density analysis on sparse data by Gaussian density function and Euclidean distance,and finally realizes the detection on the functional connectivity of reduced fMRI data.The task-related data experiment gets the following results:SDAPC produces a fine ROC curve for single subject while running about three times faster than SAPC.SDAPC and SAPC both get better ROC curves for multiple subjects than single subject.The resting-state data experiment leads to the further finding that SDAPC can successfully identify nine resting-state networks.%利用仿射聚类(APC)方法分析数据量庞大的功能磁共振成像(fMRI)数据时,在时间复杂度、数据存储和聚类效果等方面存在局限性.为此,提出一种融合稀疏仿射传播聚类(SAPC)和相似度矩阵约减的新方法(SDAPC).对fMRI数据进行稀疏逼近后,结合高斯密度函数和欧式距离对稀疏数据进行密度分析,完成约减后fMRI数据的功能连通性检测.任务态数据实验结果表明,对于单被试,SDAPC的ROC曲线与SAPC接近,但运行速度比SAPC提高了约3倍;对于多被试,SDAPC和SAPC的ROC曲线效果均优于其单被试的ROC曲线.静息态数据实验结果进一步表明,SDAPC能成功提取出9个静息态脑网络.