Detecting discriminative functional MRI activation patterns using space filling curves

We propose a novel approach for detecting discriminative patterns of functional MRI (fMRI) activation that are associated with non-spatial clinical variables (e.g. disease). The main idea is to map the 3D volumes to 1D following the traversal of the Hilbert space-filling curve, which has been shown to exhibit optimality in preserving the locality of the voxels after the domain transformation. We apply statistical tests of significance on groups of points, i.e., segments of the transformed domain, to detect discriminative patterns. To discover discriminative areas, we project these patterns to the initial 3D space by following the inverse mapping of the transformation. As a case study, we analyze an fMRI data set obtained from a study that explores neuroanatomical correlates of semantic processing in Alzheimer's disease. We seek to discover activation regions that discriminate controls from patients. We evaluate the results by presenting classification experiments that utilize information extracted from these regions. The discovered areas identified through back-projection are within the medial temporal lobe being consistent with prior findings. The overall classification accuracy ranged from 81% up to 100% for certain experimental settings. The proposed approach has great potential for elucidating structure-function relationships and can be valuable to human brain mapping.