An fMRI segmentation method under markov random fields for brain decoding

Aksan, Emre
In this study, a specially tailored segmentation method for partitioning the fMRI data into a set of "homogenous" regions with respect to a predefined cost function is proposed. The proposed method, referred as f-MRF, employs univariate and multivariate fMRI data analysis techniques under Markov Random Fields to estimate the segments by resolving a mixture density. The univariate approach helps identifying activation pattern of a voxel independently from other voxels. In order to capture local interactions among the voxels, pairwise functional similarity is used across a neighborhood. By incorporating both the unary and pairwise features of the voxels into the MRF energy function, we achieve to cluster the voxels in the brain into functionally homogeneous and spatially coherent segments. In the proposed study, voxel space is modeled with a Gaussian Mixture Model (GMM) over the univariate activation patterns, while the cluster labels are modeled as discrete Markov Random Field over the pairwise interactions. For estimation of the latent cluster labels, a two-step iterative approach is followed. Accordingly, given the current estimate of the model parameters, cluster labels are computed by using a graph-cut algorithm. In turn, the cluster labels are used to estimate the model parameters by employing maximum likelihood estimation (MLE). The final labeling result generally consists of few large clusters involving the non-activated voxels, and isolates the activated voxels into smaller-sized clusters. By partitioning the voxel space into functionally homogeneous parcels, we expect to increase representative power of the data. Thus, we propose using the f-MRF segmentation in brain decoding tasks where the segments are employed in voxel selection or feature extraction steps. In the experiments that are conducted on the real fMRI data of visual object recognition, f-MRF outperforms compared segmentation methods. Moreover, the results indicate that f-MRF has potential to boost the performance in brain decoding studies.
Citation Formats
E. Aksan, “An fMRI segmentation method under markov random fields for brain decoding,” M.S. - Master of Science, Middle East Technical University, 2015.