Anatomic Validation of Spatial Normalization Methods for PET

Spatial normalization methods, which are indispensable for intersubject analysis in current PET studies, have been improved in many aspects. These methods have not necessarily been evaluated as anatomic normalization methods because PET images are functional images. However, in view of the close relation between brain function and morphology, it is very intriguing how precisely normalized brains coincide with each other. In this report, the anatomic precision of spatial normalization is validated with three different methods. Four PET centers in Japan participated in this study. In each center, six normal subjects were recruited for both H2(15)O-PET and high-resolution MRI studies. Variations in the location of the anterior commissure (AC) and size and contours of the brain and the courses of major sulci were measured in spatially normalized MR images for each method. Spatial normalization was performed as follows. (a) Linear: The AC-posterior commissure and midsagittal plane were identified on MRI and the size of the brain was adjusted to the Talairach space in each axis using linear parameters. (b) Human brain atlas (HBA): Atlas structures were manually adjusted to MRI to determine linear and nonlinear transformation parameters and then MRI was transformed with the inverse of these parameters. (c) Statistical parametric mapping (SPM) 95: PET images were transformed into the template PET image with linear and nonlinear parameters in a least-squares manner. Then, coregistered MR images were transformed with the same parameters used for the PET transformation. The AC was well registered in all methods. The size of the brain normalized with SPM95 varied to a greater extent than with other approaches. Larger variance in contours was observed with the linear method. Only SPM95 showed significant superiority to the linear method when the courses of major sulci were compared. The results of this study indicate that SPM95 is as effective a spatial normalization as HBA, although it does not use anatomic images. Large variance in structures other than the AC and size of the brain in the linear method suggests the necessity of nonlinear transformations for effective spatial normalization. Operator dependency of HBA also must be considered.