Abstract:
|
Standard tests in brain image analyses rely on an assumption of stationary smoothness, or homogeneous spatial correlation structure, which is especially crucial in cluster size tests. If this assumption fails, the tests become conservative in rough regions and liberal in smooth regions. This is a particular concern for images from voxel-based morphometry whose smoothness varies dramatically from cortical to subcortical regions. To correct biases under such circumstances, we propose a cluster size permutation test which measures cluster volume as sums of resels (resolution elements), which are inversely related to the smoothness of images. The maximum cluster size in units of resel was used as a test statistic for a permutation test of 1000 permutations comparing fMRI data of six schizophrenics and eight normals, and the resulting p-values were compared to a standard permutation test with a stationary smoothness assumption. Though our test is more conservative due to the variability in resel estimation, p-values were reduced in relatively rough regions and increased in relatively smooth regions, suggesting our test's ability to adjust significance according to local smoothness.
|