MCI的rs-fMRI功能性连接的特征选择与压缩

Feature Selection and Compression of rs-fMRI Functional ConnectioninMCI

轻度认知障碍(mild cognitive impairment,MCI)的诊断和及时治疗对阿兹海默症(alzheimer’s disease,AD)患者提供早期预警信号具有临床意义.通过神经影像学技术和机器学习(machine learning,ML)对MCI进行辅助诊断的方法性能主要依赖于筛选可表达组间显著性差异的特征,而目前常用皮尔逊相关法表示脑区连通性并将其直接作为分类器的输入特征,通常这些特征包含冗余信息且易造成维度诅咒的问题.针对该问题,提出特征选择和特征压缩相结合的方法筛选重要特征,首先对rs-fMRI计算动态功能连接(dynamic functional connectivity,DFC),其次利用最小类内距离准则筛选重要的特征,然后对筛选后的特征进行最小二乘(least square,LS)线性拟合压缩数据,最后将得到的拟合系数作为分类器输入特征.实验结果表明,特征压缩与特征选择结合的算法获得的分类精度可达76%,比未经特征处理的分类准确率提高了大约8%,表明该方法能有效提高MCI分类准确率,具有一定的生物学意义.

The diagnosis and timely treatment of mild cognitive impairment(MCI)has clinical significance in providing early warning signals for Alzheimer s Disease.The performance of the methods for auxiliary diagnosis of MCI by neuroimaging technology and machine learning(ML)is mainly depend on how to screenfor features that can express significant differences between groups.The commonly used Pearson correlation method represents the brain functional connection(FC)and take it as the input features of the classifier.Usually,A large number of features that contain redundant information can create a dimension curse problem.Thus,proposing a method combining feature selection and feature compression to get important features to solve the problem.Firstly,Sliding window technology is used to perform dynamic functional connectivity(DFC).Secondly,the minimum in-class distance criterion is used to get important features.Then the selected features are compressed by least square(LS).Finally,the fitting coefficient obtained by LS is used as the latest feature to achieve classification.The experimental results show that the classification accuracy of the algorithm combining feature compression and feature selection can achieve 76%,which is about 8%higher than the traditional method.It can effectively improve the classification accuracy of MCI,which has certain biological significance.