The interpretation of functional brain images is often hampered by the presence of noise. This problem is most commonly solved by using a statistical method and only considering signals that are unlikely to occur by chance. The method used should be specific and sensitive, specific because only true signals are of interest and sensitive because this will enable more information to be extracted from each experiment. Here we present a modification of the cluster analysis proposed by Roland et al. (Human Brain Mapping 1: 3-19, 1993). A covariance model is used to test hypotheses for each voxel. The generated statistical images are searched for the largest clusters. From the same data set noise images are generated. For each of these noise images the autocorrelation function is estimated. These estimates are subsequently used to generate simulated noise images, from which a distribution of cluster sizes is derived. The derived distribution is used to estimate probabilities for the clusters detected in the statistical images generated by testing the hypothesis. This presented method is shown to be specific and is further compared with SPM96 and the nonparametric method of Holmes et al. (J. Cereb. Blood Flow Metab. 16: 7-22, 1996).
Copyright 1998 Academic Press.