Multifactor Analysis for fMRI Brain Image Classification by Subject and Motor Task

FMRI brain images are generated by the variation of multiple factors, such as subject, motor task, and time frame. Just as this example demonstrates, in image analysis, much work has been aimed at analyzing a set of images generated by variation of multiple factors. To perform image analysis successfully, it is often necessary to model multiple factor frameworks found in image sets. One leading method for multifactor analysis is Multilinear Principal Component Analysis (MPCA), also called Tensorfaces, based on Higher-Order Singular Value Decomposition in tensor algebra. While traditional dimension reduction methods use a single low-dimensional vector to represent an original high-dimensional sample vector, MPCA uses multiple low-dimensional vectors associated with multiple factors. This project extends Multifactor Kernel PCA (MKPCA), a kernel-based version of MPCA, based on kernels that consist of multiple sub-kernels. Based on multifactor analysis provided by MKPCA, in this project, we conducted experiments using a set of 4D spatiotemporal fMRI brain images with two factors: 14 subjects and 4 motor tasks. Subject parameters and motor-task parameters, obtained by decomposing the influences of the two factors, were used for classifying fMRI images by subject and motor task, respectively. The classification accuracy obtained by our experiments demonstrates the advantages of our proposed method over current leading methods for multifactor analysis. In particular, MKPCA outperformed MPCA, as MKPCA was able to provide more reliable analysis for nonlinear structures in real-world data sets.