Abstract
Independent component analysis (ICA) is widely used in the field of functional neuroimaging to decompose data into spatio-temporal patterns of co-activation. In particular, ICA has found wide usage in the analysis of resting state fMRI (rs-fMRI) data. Recently, a number of large-scale data sets have become publicly available that consist of rs-fMRI scans from thousands of subjects. As a result, efficient ICA algorithms that scale well to the increased number of subjects are required. To address this problem, we propose a two-stage likelihood-based algorithm for performing group ICA, which we denote Parallel Group Independent Component Analysis (PGICA). By utilizing the sequential nature of the algorithm and parallel computing techniques, we are able to efficiently analyze data sets from large numbers of subjects. We illustrate the efficacy of PGICA, which has been implemented in R and is freely available through the Comprehensive R Archive Network, through simulation studies and application to rs-fMRI data from two large multi-subject data sets, consisting of 301 and 779 subjects respectively.
Highlights
Independent component analysis (ICA) is a blind source separation technique [1] that assumes the observed signals are linear mixings of independent underlying sources
The method of [14] is theoretically scalable in terms of its memory requirements, the approach requires the serial calculation of gradients to optimize parameter estimation, which can be very slow for high-dimensional data to the point where it would still be practically infeasible in terms of computation time
Computing the gradients for different subjects in parallel could potentially speed up the algorithm dramatically, provided the cost is much lower than the necessary data transfer time
Summary
Independent component analysis (ICA) is a blind source separation technique [1] that assumes the observed signals are linear mixings of independent underlying sources. Following the estimation of group-level ICs, a wide variety of methods can be used to reconstruct subject-specific independent components, such as GICA 1, GICA 2, GICA 3, dual regression and Group Information Guided ICA (GIG-ICA) Both dual regression and GIG-ICA have great scalability [5,6,7]. Scalable PCA/SVD algorithms are required to handle large data efficiently in group ICA. The proposed Parallel Group Independent Component Analysis (PGICA) is different from fastICA and JADE in that the algorithm is likelihood-based and uses maximum likelihood estimation (MLE) for parameter estimation. Current group ICA algorithms have limited power for scaling to analyze large data sets, especially in the field of resting state fMRI analysis because they require data to first be concatenated across subjects and reduced via PCA prior to estimation of group-level independent components. The first is a collection of 301 adults, while the second is a set of 779 fMRI scans, consisting of 379 with autism spectrum disorder (ASD) and 400 typically developing controls
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
