Threshold-free cluster enhancement (TFCE) : improving power and stability of cluster size inference in brain imaging

Cluster size inference, or tests based on the spatial extent of brain imaging signals, is a widely used approach in Functional Magnetic Resonance Imaging (fMRI). Clusters are formed by first applying a cluster-forming threshold, and then the significance is assessed based on the volume of the cluster. While this has been shown to be more powerful than 'voxel-wise' inference (inference based on statistic image magnitude at each volume element) it depends critically on the cluster-forming threshold, u. Small changes to u can split or join clusters and change the significance in unexpected way. We have created a new method that merges clustering information over all possible u's, using permutation resampling methods to obtain valid P-values. While elminating one nuisance parameter (u), our method introduces two new ones. I will show how we used (parametric) random field theory to fin