Influence of ANOVA design and anatomical standardization on statistical mapping for PET activation

We have created images of z value, error, and variation components for a PET activation study using various ANOVA designs and anatomical standardization methods. Data were acquired in four PET centers. In each center, CBF was measured on six normal male subjects under resting and covert verb generation, three times for each. The images were anatomically standardized with LINEAR transformation, SPM (Ver. 95), HBA (Karolinska/Tohoku), or MICHIGAN (Minoshima). ANOVA was performed pixel by pixel to compute t (and z) for the task main effect (Verb vs Rest) in four different designs: (i) two way (subject and task) (2W), (ii) two-way with interaction (2WI), (iii) subject considered a random factor (2WI-MX), and (iv) three-way (subject, task, and replication) (3W). A large area extending from the Broca to the left premotor cortex was activated. The localization of the highest peak depended both on the anatomical standardization and on the ANOVA design, the variation ranging 3-4 cm. Smoothing reduced the variation while erasing possible subfoci. The z images of 2W, 2WI, and 3W looked alike, whereas 2WI-MX presented lower peak z values. SPM tended to present higher z values than the other methods. The error was high in the gray and low in the white matter. The root mean square for the subject effect was high on the border of gray matter especially in LINEAR and HBA, revealing intersubject mismatch in the gray matter distribution. The root mean square for the subject-by-task interaction effect revealed individual variation in activation.