fMRI Activation Detection by MultiScale Hidden Markov Model

This paper considers detection of functional magnetic resonance images (fMRIs), that is, to decide active and nonactive regions of human brain from fMRIs. A novel two-step approach is put forward that incorporates spatial correlation information and is amenable to analysis and optimization. First, a new multi-scale image segmentation algorithm is proposed to decompose the correlation image into several different regions, each of which is of homogeneous statistical behavior. Second, each region will be classified independently as active or inactive using existing pixel-wise test methods. The image segmentation consists of two procedures: edge detection followed by label estimation. To deduce the presence or absence of an edge from continuous data, two fundamental assumptions of our algorithm are 1) each wavelet coefficient is described by a 2-state Gaussian Mixture Model (GMM); 2) across scales, each state is caused by its parent state, hence the name of multiscale hidden Markov model (MHMM). The states of Markov chain are unknown ("hidden") and represent the presence (state 1) or absence (state 0) of edges. Using this interpretation, the edge detection problem boils down to the posterior state estimation given obervation.