Abstract
Hypothesis-driven methods (such as general linear model analysis) have been widely adopted for functional magnetic resonance imaging (fMRI) analysis. However, assumptions on hemodynamic response function (HRF) models are always controversial, and a precise relationship between the external stimulus and the brain reaction is essential. On the contrary, data-driven methods (e.g. principle component analysis, independent component analysis, and clustering analysis) are free from these problems, due to their model-free property. The major difference between component analyses and clustering analysis is: original data is decomposed into several components in former methods, while the latter separates data into groups via similarities. The concept of components may be intuitive but questionable because of the strong assumptions on the linearity of components, whereas clustering analysis bears the advantage of the fewest assumptions. Furthermore, in all data-driven methods dimension reduction strategies must be applied to overcome the high computational complexity, and feature extraction preserves key information in spite of dimension reduction. Therefore, we proposed to use signal corresponding features as well as noise features for clustering. Since autocorrelation function (ACF) is a common tool to estimate noise, ACF features can thus be used to form the feature space before applying any clustering algorithm.