|Authors||Whitcher, B., Schwarz, A.J., Barjat, H., Smart, S.C., Grundy, R.I. and James, M.F.|
MRI time series experiments produce a wealth of information contained in two or three spatial dimensions that evolve over time. Such experiments can, for example, localize brain response to pharmacological stimuli, but frequently the spatiotemporal characteristics of the cerebral response are unknown a priori and variable, and thus difficult to evaluate using hypothesis-based methods alone. Here we used features in the temporal dimension to group voxels with similar time courses based on a nonparametric discrete wavelet transform (DWT) representation of each time course. Applying the DWT to each voxel decomposes its temporal information into coefficients associated with both time and scale. Discarding scales in the DWT that are associated with high-frequency oscillations (noise) provided a straight-forward data reduction step and decreased the computational burden. Optimization-based clustering was then applied to the remaining wavelet coefficients in order to produce a finite number of voxel clusters. This wavelet-based cluster analysis (WCA) was evaluated using two representative classes of MRI neuroimaging experiments. In perfusion-weighted MRI, following occlusion of the middle cerebral artery (MCAO), WCA differentiated healthy tissue and different regions within the ischemic hemisphere. Following an acute cocaine challenge, WCA localized subtle differences in the pharmacokinetic profile of the cerebral response. We conclude that WCA provides a robust method for blind analysis of time series image data.