Implementation and Application of Principal Component Analysis on Functional Neuroimaging Data

Abstract

Recent interest has arisen regarding the application of principal component analysis (PCA)-style methods for the analysis of large neuroimaging data sets. However, variation between different implementation of these techniques has resulted in some confusion regarding the uniqueness of these approaches. <p>In the present article, we attempt to provide a more unified insight into the use of PCA as a useful method of analyzing brain image data contrasted between experimental conditions. We expand on the general approach by evaluating the use of permutation tests as a means of assessing whether a given solution, as a whole, exposes significant effects of the task difference. This approach may have advantages over more simplistic methods for evaluating PCA results and does not require extensive or unrealistic statistical assumptions made by conventional procedures. <p>Furthermore, we also evaluate the use of axes rotation on the interpretability of patterns of PCA results. Finally, we comment on the variety of PCA-style techniques in the neuroimaging literature that are motivated largely by the kind of research question being asked and note how these seemingly disparate approaches differ in how the data is preprocessed not in the fundamentals of the underlying mathematical model.