Activation detection in event-related fMRI through clustering ofwavelet distributions

We propose a new method for the detection of activated voxels in event-related BOLD fMRI data. We model the statistics of the wavelet histograms derived from each voxel time series independently through a generalized Gaussian distribution (GGD). We perform k-means clustering of the GGDs characterizing the voxel data in a synthetic data set, using the symmetrized Kullback-Leibler divergence (KLD) as a similarity measure. We compare our technique with GLM modeling and with another clustering method for activation detection that directly uses the wavelet coefficients as features. Our method is shown to be considerably more stable against realistic hemodynamic variability.