An information-based clustering approach for fMRI activation detection

Most clustering algorithms in fMRI analysis implicitly require some nontrivial assumption on data structure. Due to arbitrary distribution of fMRI time series in the temporal domain, such analysis may mislead and limit the detector's performance. In this work, the authors exploited the application of an information-based clustering algorithm (Iclust) which could avoid these assumptions and provide many other benefits, such as no cluster shape restriction, no need of a prior definition about similarity measure, and the ability of capturing both linear and nonlinear dependence. Results from both artificial and real fMRI data indicated that the proposed framework could achieve better spatiotemporal accuracy, and enabled the exploration of fine functional distinction of the human visual system in accordance with its well-known anatomy organizations.