基于体素的fMRI数据分类研究及其应用  

RESEARCH AND APPLICATION OF VOXEL-BASED fMRI DATA CLASSIFICATION

作　　者：张兵[1] 董云云[1] 邓红霞[1] 李海芳[1] 

机构地区：[1]太原理工大学计算机科学与技术学院,山西太原030024

出　　处：《计算机应用与软件》2015年第2期138-142,共5页Computer Applications and Software

基　　金：国家自然科学基金项目(61070077;61170136);山西省自然科学基金项目(2010011020-2)

摘　　要：使用机器学习方法分类f MRI(functional magnetic resonance imaging)数据已经逐渐广泛被应用到探索大脑认知的研究中。在探索人脑视觉区域对颜色特征和形状特征的捆绑图像认知研究中,使用血氧含量水平BOLD(blood oxygen level dependent)最大值、BOLD变化累计值作为特征值训练SVM分类器,使用BOLD变化时间序列方差及均值组合作为特征值训练多个SVM弱分类器,并使用Adaboost算法将多个SVM分类器集成到一起构造集成分类器,以此来判断人正在观察的图像的类型。实验结果表明,使用BOLD时间序列方差及均值组合作为特征构造的集成分类器分类正确率较高,对比不同视觉区域对特征捆绑任务识别正确率,发现V3区对图像复杂度的改变比较敏感,与特征捆绑的任务联系比较紧密。该方法可以应用到脑机接口BCI(brain computer interface)等领域。Using machine learning technology to classify fMRI （functional magnetic resonance imaging） data has been gradually and widely applied in the studies of exploring brain cognition. In the research of exploring the cognition of human brain visual area on the images biding the colour and shape features, we use the maximum value of blood oxygen level dependent （BOLD） and the cumulative change value of BOLD as the eigenvalues to train SVM classifier, and use the combination of the variance and the mean of BOLD changes time series as the cigenvalue to train multiple SVM weak classifiers, furthermore, the Adaboost algorithm is employed to integrate multiple SVM classifiers to construct an ensemble classifier. With these it is able to determine the type of the images men are observing. Experimental result demonstrates that the ensemble classifier constructed using the combination of variance and mean of BOLD time series as the features has higher classification accuracy. Comparing the recognition accuracy of different visual areas on features binding task, it is found that the V3 area is sensitive to changes in image complexity and has closer links with the features-bound tasks. This method can be used to the field of brain computer interface （BCI）, etc.

关 键 词：SVM分类 FMRI BOLD变化模式 视觉区域 特征捆绑 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]