Clustering of fMRI Data Using Affinity Propagation

Clustering methods are commonly used for fMRI (functional Magnetic Resonance Imaging) data analysis. Based on an effective clustering algorithm called Affinity Propagation (AP) and a new defined similarity measure, we present a method for detecting activated brain regions. In the proposed method, autocovariance function values and the Euclidean distance metric of time series are firstly calculated and combined into a new similarity measure, then the AP algorithm with the measure is carried out on all time series of data, and at last regions with which their cross-correlation coefficients are greater than a threshold are taken as activations. Without setting the number of clusters in advance, our method is especially appropriate for the analysis of fMRI data collected with a periodic experimental paradigm. The validity of the proposed method is illustrated by experiments on a simulated dataset and a benchmark dataset. It can detect all activated regions in the simulated dataset accurately, and its error rate is smaller than that of K-means. On the benchmark dataset, the result is very similar to SPM.