Abstract

Threshold-free Cluster Enhancement (TFCE) was introduced as a method to overcome the limitations of using cluster size inference for detecting effects in neuroimaging statistic maps. TFCE has seen wide application (with the original paper paper having over 6000 citations) likely due to its high sensitivity, based on an image providing a voxel-wise representation of "cluster-like local spatial support". In this talk I will demonstrate that TFCE's sensitivity comes at the cost of spatial specificity. Specifically, I will show that TFCE does not provide voxel-wise or even cluster-wise error control. Most troubling, this means that an isolated TFCE detection can imply true signals almost anywhere in the brain. Embedding TFCE within a closed testing framework, I propose Localized Cluster Enhancement (LCE) which addresses these problems. I will show that LCE can be used to detect activity within anatomical or data-driven regions of interest while correctly controlling false positive rates.