Overcoming the ill-balanced data problem in functional MRI clustering analysis

In functional magnetic resonance imaging (fMRI) data, activated voxels are usually very small in number and are embedded in a mass of inactive voxels. For clustering analysis, this situation generates an ill-balanced data problem among different classes of voxels. In this paper we propose a novel method to overcome the ill-balanced data problem, by reducing the number of voxels to be processed by the clustering algorithm. We divide the functional data into small overlapping regions and decide the presence or absence of functional activity in a region, on the basis of condition number of a matrix constructed from the feature vectors of the voxel in that region. Only the regions that are potentially active are retained for clustering analysis. Results are presented for both simulated and real data and advocate that the proposed method effectively solves the ill-balanced data problem.