Data-Driven Optimization of Brain SPECT Processing for Voxel-Based Statistical Analysis

Brain SPECT processing methods were assessed by reprocessing a SPECT study of aging in brain perfusion that revealed both localized preservation and losses. Each analysis was characterized by the Z statistic for linear regression of SPECT versus age in six locations: three with preservation, three with accelerated loss. We investigated the effect of reconstruction (FBP or OSEM), scatter subtraction (SS), removal of facial and scalp activity, global scaling to the whole-brain and cerebellum, and spatial normalization with both SPM99 and SPM2. Facial activity was edited manually. Scalp activity was removed using a generic weighting image after an initial affine spatial normalization. Global scaling was performed with SPMs "proportional scaling," or a scalp-free mean applied either as a nuisance variable (ANCOVA) or to prescale the images. Stronger statistics resulted from FBP with SS, removal of facial and scalp activity and nonlinear spatial normalization. Proportional scaling was biased and yielded very strong SPECT preservation but no significant SPECT loss. The scalp-free mean as a nuisance variable yielded both significant SPECT preservation and losses, but statistics were strongest when it was used to prescale the images. These scaling effects arose from age-dependence in the global means. Scaling to the cerebellum mean yielded weaker results than the whole brain. SPM99 yielded similar results to SPM2. Scaling of brain SPECT images should utilize a scalp-free mean and age should be included as a covariate in regional SPECT analysis