On two methods of statistical image analysis

Abstract

The computerized brain atlas (CBA) and statistical parametric mapping (SPM) are two procedures for voxel-based statistical evaluation of PET activation studies. Each includes spatial standardization of image volumes, computation of a statistic, and evaluation of its significance. In addition, smoothing and correcting for differences of global means are commonly performed in SPM before statistical analysis. We report a comparison of methods in an analysis of regional cerebral blood flow (rCBF) in 10 human volunteers and 10 simulated activations. For the human studies, CBA or linear SPM standarization methods were followed by smoothing and computation of a statistic with the paired t-test of CBA or general linear model of SPM. No standardization, linear, and nonlinear SPM standardization were applied to the simulations. Significance of the statistic was evaluated using the cluster-size method common to SPM and CBA. SPM employs the theory of Gaussian random fields to estimate the cluster size distributions; simulations described in the Appendix provided empirical distributions derived from t-maps. The quantities evaluated were number and size of functional regions (FRs), maximum statistic, average resting rCBF, and percentage change. For the simulations, the efficiency of signal detection and rate of false positives could be evaluated as well as the distributions of statistics and cluster size in the absence of signal. The similarity of the results yielded by similar methods of analysis for the human studies and the simulated activations substantiates the robustness of the methods for selecting functional regions. However, the analysis of simulated activations demonstrated that quantitative evaluation of significance of a functional region encounters important obstacles at every stage of the analysis.