基于fMRI的视觉信息解码研究进展

摘要

从视觉信息解码的层次、数学模型和数据预处理的关键步骤对基于功能磁共振成像数据的神经解码技术进行综述.根据解码的最终目标,将视觉信息的神经解码分为类别(如不同客体类别)与内容解码(如重构被试看到的字母刺激)两个层次.分析了神经解码数学模型中的多体元模式分析技术,将其分为线性与非线性模型两类,其中以线性支持向量机、稀疏性算法应用较多.比较了数据预处理中平滑处理、特征选择与特征空间转换对解码模型的影响,结果表明,结合非平滑处理的基于信息的多变量特征选择方法对解码影响较大.未来的研究重点集中在视觉信息的内容解码,并由二维图像扩展到对三维图像信息的解码.%This paper reviews three essential aspects in visual information decoding of fMRI data,including the level of decoding,mathematical model and data pre-processing.The visual decoding is divided into two levels:classification (e.g.,prediction of the predefined categories of objects) and content decoding (e.g.,the reconstruction of letters).The multi-voxel pattern analysis (MVPA) techniques are studied,which can be categorized into linear and nonlinear methods;and the linear support vector machine and sparse algorithms are more commonly used.The different impacts of smooth pre-processing,feature selection and feature space transformation on the decoding model have been analyzed,and the results indicate that the information-based feature selection method on unsmooth data achieves better results.In the future,this research will focus on visual information reconstruction,and extend from two-dimensional to three-dimensional visual stimuli.