Statistical parametric mapping of brain morphology: sensitivity is dramatically increased by using brain-extracted images as inputs.

A major attraction of voxel-based morphometry (VBM) is that it allows researchers to explore large datasets with minimal human intervention. However, the validity and sensitivity of the Statistical Parametric Mapping (SPM2) approach to VBM are the subject of considerable debate. We visually inspected the SPM2 gray matter segmentations for 101 research participants and found a gross inclusion of non-brain tissue surrounding the entire brain as gray matter in five subjects and focal areas bordering the brain in which non-brain tissue was classified as gray matter in many other subjects. We also found many areas in which the cortical gray matter was incorrectly excluded from the segmentation of the brain. The major source of these errors was the misregistration of individual brain images with the reference T1-weighted brain template. These errors could be eliminated if SPM2 operated on images from which non-brain tissues (scalp, skull, and meninges) are removed (brain-extracted images). We developed a modified SPM2 processing pipeline that used brain-extracted images as inputs to test this hypothesis. We describe the modifications to the SPM2 pipeline that allow analysis of brain-extracted inputs. Using brain-extracted inputs eliminated the non-brain matter inclusions and the cortical gray matter exclusions noted above, reducing the residual mean square errors (RMSEs, the error term of the SPM2 statistical analyses) by over 30%. We show how this reduction in the RMSEs profoundly affects power analyses. SPM2 analyses of brain-extracted images may require sample sizes only half as great as analyses of non-brain-extracted images.