Equitable Thresholding and Clustering

Abstract

This paper describes a hybrid method to threshold FMRI group statistical maps derived from voxelwise second-level statistical analyses. The proposed \"Equitable Thresholding and Clustering\" (ETAC) approach seeks to reduce the dependence of clustering results on arbitrary parameter values by using multiple sub-tests, each equivalent to a standard FMRI clustering analysis, to make decisions about which groups of voxels are potentially significant. The union of these sub-test results decides which voxels are accepted. The approach adjusts the cluster-thresholding parameter of each sub-test in an equitable way, so that the individual false positive rates (FPRs) are balanced across sub-tests to achieve a desired final FPR (e.g., 5%). ETAC utilizes resampling methods to estimate the FPR, and thus does not rely on parametric assumptions about the spatial correlation of FMRI noise. The approach was validated with pseudo-task timings in resting state brain data. Additionally, a task FMRI data collection was used to compare ETACs true positive detection power vs. a standard cluster detection method, demonstrating that ETAC is able to detect true results and control false positives while reducing reliance on arbitrary analysis parameters.