Statistical Segmentation of fMRI Activations Using Contextual Clustering

Abstract

A central problem in the analysis of functional magnetic resonance imaging (fMRI) data is the reliable detection and segmentation of activated areas. Often this goal is achieved by computing a statistical parametric map (SPM) and thresholding it. Cluster-size thresholds are also used. A new contextual segmentation method based on clustering is presented in this paper. If the SPM value of a voxel, adjusted with neighborhood information, differs from the expected non-activation value more than a specified decision value, the contextual clustering algorithm classifies the voxel to the activation class, otherwise to the non-activation class. The voxel-wise thresholding, cluster-size thresholding and contextual clustering are compared using fixed overall specificity. Generally, the contextual clustering detects activations with higher probability than the voxel-wise thresholding. Unlike cluster-size thresholding, contextual clustering is able to detect extremely small area activations, too. Moreover, the results show that the contextual clustering has good segmentation accuracy, voxel-wise specificity and robustness against spatial autocorrelations in the noise term.