Sariya Yogesh Kumar, Anand R S
Instrumentation and Signal Processing Group, Department of Electrical Engineering, Indian Institute of Technology Roorkee, Roorkee-247667, India.
J Integr Neurosci. 2017;16(2):157-175. doi: 10.3233/JIN-170006.
Independent component analysis, a data-driven analysis method, has found significant applications in task-based as well as resting state fMRI studies. There are numbers of independent component analysis algorithms available, but only a few of them have been used frequently so far for fMRI images. With a view that algorithms that are overlooked may outperform the most opted, a comparative study is taken up in this paper to analyze their abilities for the purpose of synthesis of fMRI images. In this paper, ten independent component algorithms: Fast ICA, INFOMAX, SIMBEC, JADE, ERICA, EVD, RADICAL, ICA-EBM, ERBM, and COMBI are compared. Their separation abilities are adjudged on both, synthetic and real fMRI images. Performance to decompose synthetic fMRI images is being monitored on the basis of spatial correlation coefficients, time elapsed to extract independent components and the visual appearance of independent components. Ranking of their performances on task-based real fMRI images are based on the closeness of time courses of identified independent components with model time course and the closeness of spatial maps of components with spatial templates while their competencies for resting state fMRI data are analyzed by examining how distinctly they decompose the data into the most consistent resting state networks. Sum of mutual information between all the permutations of decomposed components of resting state fMRI data are also calculated.
独立成分分析作为一种数据驱动的分析方法，已在基于任务的以及静息态功能磁共振成像（fMRI）研究中得到了广泛应用。现有多种独立成分分析算法，但到目前为止，其中只有少数几种经常用于fMRI图像。鉴于那些被忽视的算法可能比最常选用的算法表现更优，本文开展了一项比较研究，以分析它们合成fMRI图像的能力。本文比较了十种独立成分算法：快速独立成分分析（Fast ICA）、信息最大化算法（INFOMAX）、SIMBEC算法、联合近似对角化特征矩阵算法（JADE）、ERICA算法、特征值分解算法（EVD）、RADICAL算法、基于期望最大化的独立成分分析算法（ICA-EBM）、基于期望最大化的拉普拉斯算法（ERBM）和COMBI算法。在合成fMRI图像和真实fMRI图像上评判它们的分离能力。基于空间相关系数、提取独立成分所花费的时间以及独立成分的视觉外观来监测合成fMRI图像的分解性能。它们在基于任务的真实fMRI图像上的性能排名是基于所识别的独立成分的时间历程与模型时间历程的接近程度以及成分的空间图谱与空间模板的接近程度，而通过检查它们将数据分解为最一致的静息态网络的清晰度来分析它们处理静息态fMRI数据的能力。还计算了静息态fMRI数据分解成分的所有排列之间的互信息总和。
