Validating and improving cluster size inference in brain image analysis.

Abstract

A typical brain image data set consists of a set of 3D images, each of which is composed of tens of thousands of intensity measurements called voxels. In a typical analysis of such data, a linear model is fitted at each voxel and an image of test statistics is created. Cluster size inference is one approach of assessing such statistic images, in which high intensity voxels are grouped into clusters and spatial extent of such clusters is used to test the null hypothesis. There are different methods for cluster size inference, a random field theory (RFT) based approach and a permutation test, to name two. Though widely used, these methods have not been extensively evaluated in realistic settings, where some of the assumptions of the tests may not hold. Violations of stationarity assumption, or uniform smoothness, are particularly problematic, since clusters tend to be large in smooth areas and small in rough areas of the image. Furthermore, since cluster size inference only uses cluster size and ignores intensity information, it may be unable to detect high intensity peaked signals. To address these issues, we assess properties of these methods by Monte-Carlo simulations under various settings. We propose a permutation cluster size test under non-stationarity, using local smoothness information to adjust sensitivity. We also propose methods that combine voxel intensities and cluster size to yield improved voxel-cluster size inference, sensitive to both spatially extended signals and high intensity signals.