Global spatial normalization transforms a brain image so that its principal global spatial features (position, orientation and dimensions) match those of a standard or atlas brain, supporting consistent analysis and referencing of brain locations. The convex hull (CH), derived from the brain's surface, was selected as the basis for automating and standardizing global spatial normalization. The accuracy and precision of CH global spatial normalization of PET and MR brain images were evaluated in normal human subjects.
Methods: Software was developed to extract CHs of brain surfaces from tomographic brain images. Pelizzari's hat-to-head least-square-error surface-fitting method was modified to fit individual CHs (hats) to a template CH (head) and calculate a nine-parameter coordinate transformation to perform spatial normalization. A template CH was refined using MR images from 12 subjects to optimize global spatial feature conformance to the 1988 Talairach Atlas brain. The template was tested in 12 additional subjects. Three major performance characteristics were evaluated: (a) quality of spatial normalization with anatomical MR images, (b) optimal threshold for PET and (c) quality of spatial normalization for functional PET images.
Results: As a surface model of the human brain, the CH was shown to be highly consistent across subjects and imaging modalities. In MR images (n = 24), mean errors for anterior and posterior commissures generally were <1 mm, with SDs < 1.5 mm. Mean brain-dimension errors generally were <1.3 mm, and bounding limits were within 1-2 mm of the Talairach Atlas values. The optimal threshold for defining brain boundaries in both 18F-fluorodeoxyglucose (n = 8) and 15O-water (n = 12) PET images was 40% of the brain maximum value. The accuracy of global spatial normalization of PET images was shown to be similar to that of MR images.
Conclusion: The global features of CH-spatially normalized brain images (position, orientation and size) were consistently transformed to match the Talairach Atlas in both MR and PET images. The CH method supports intermodality and intersubject global spatial normalization of tomographic brain images.