fMRI time-series clustering using a mixture of mixtures of Student's-t and Rayleigh distributions

In this paper, a new Markov random field-based mixture model, where each of its components is a mixture of Student's-t and Rayleigh distributions, is proposed for clustering fMRI time-series. By introducing the non-symmetric Rayleigh distribution, the proposed algorithm has flexibility to fit various types of observed time-series. Moreover, our method incorporates Markov random field so that the spatial relationships between neighboring voxels are considered, which makes the presented model more robust to noise, and that preserves more details of the clustering results compared with other symmetric distribution-based algorithms. Additionally, the expectation maximization algorithm is directly implemented to estimate the parameter set by maximizing the data log-likelihood function. The proposed framework is evaluated on real fMRI time-series, and the quantitatively compared results are demonstrated in terms of effectiveness and accuracy.