Comparison of spatial normalization procedures and their impact on functional maps

Abstract

The alignment accuracy and impact on functional maps of four spatial normalization procedures have been compared using a set of high resolution brain MRIs and functional PET volumes acquired in 20 subjects. Simple affine (AFF), fifth order polynomial warp (WRP), discrete cosine basis functions (SPM), and a movement model based on full multi grid (FMG) approaches were applied on the same dataset for warping individual volumes onto the Human Brain Atlas (HBA) template. Intersubject averaged structural volumes and tissue probability maps were compared across normalization methods and to the standard brain. Thanks to the large number of degrees of freedom of the technique, FMG was found to provide enhanced alignment accuracy as compared to the other three methods, both for the grey and white matter tissues; WRP and SPM exhibited very similar performances whereas AFF had the lowest registration accuracy. SPM, however, was found to perform better than the other methods for the intra-cerebral cerebrospinal fluid (mainly in the ventricular compartments). Limited differences in terms of activation morphology and detection sensitivity were found between low resolution functional maps (FWHM approximately 10 mm) spatially normalized with the four methods, which overlapped in 42.8% of the total activation volume. These findings suggest that the functional variability is much larger than the anatomical one and that precise alignment of anatomical features has low influence on the resulting intersubject functional maps. When increasing the spatial resolution to approximately 6 mm, however, differences in localization of activated areas appear as a consequence of the different spatial normalization procedure used, restricting the overlap of the normalized activated volumes to only 6.2%.

Figure 1

A: Original axial section of the SPM segmentation applied to the HBA MRI template. B: Morphologically processed slice. This process removes the extra cerebral CSF tissue that remains after the semi‐automated scalp edition and fills‐in the ventricles and deepness of the sulci. C: Corresponding MRI slice. D: Transverse processed sections of the segmentation of the HBA MRI template from +70 to −40 mm relative to the AC–PC plane.

Figure 2

From left to right: transverse (z = +50 mm), sagittal (x = −44 mm) and coronal (y = +20 mm) sections of the HBA MRI template (top row) and of the intersubject average volumes computed for the four normalization procedures. Coordinates are relative to the Talairach space. AFF, affine transformation (n = 20); SPM, SPM linear and nonlinear routines of the SPM package (n = 20); WRP, affine and fifth order polynomial normalization of the AIR package (n = 20); FMG, affine and nonlinear fluid transformation (n = 19).

Figure 3

Left column: Axial section (z = +50 mm) of the HBA MRI template and enlargement of the corresponding segmented grey matter volume focused on the superior frontal gyrus. Middle column: Grey matter probability maps (n = 18) computed for the four normalization procedures. These maps are restricted to voxels located inside the corresponding HBA grey matter template and denoted “Classified voxels.” Right column: Grey matter probability maps obtained for voxels located outside the corresponding HBA grey matter template and denoted “Misclassified voxels.” AFF, affine transformation; SPM, SPM linear and nonlinear routines of the SPM package; WRP, affine and fifth order polynomial normalization of the AIR package; FMG, affine and nonlinear fluid transformation.

Figure 4

Left column: Axial sections (z = +50 mm) of the HBA MRI template and of an enlargement of the corresponding segmented white matter volume focused on the superior frontal gyrus. Middle column: White matter probability maps (n = 18) computed for the four normalization procedures. These maps are restricted to the voxels located inside the corresponding HBA white matter template and denoted “Classified voxels.” Right column: White matter probability maps obtained for voxels located outside the corresponding HBA white matter template and denoted “Misclassified voxels.” AFF, affine transformation; SPM, SPM linear and nonlinear routine of the SPM package; WRP, affine and fifth order polynomial normalization of the AIR package; FMG, affine and nonlinear fluid transformation.

Figure 5

Left column: Axial sections (z = +18 mm) of the HBA MRI template and of an enlargement of the corresponding segmented CSF volume focused on the ventricles. Middle column: CSF probability maps (n = 18) computed for the four normalization procedures. These maps are restricted to the voxels located inside the corresponding HBA CSF template and denoted “Classified voxels.” Right column: CSF probability maps obtained for voxels located outside the corresponding HBA CSF template and denoted “Misclassified voxels.” AFF, affine transformation; SPM, SPM linear and nonlinear routine of the SPM package; WRP, affine and fifth order polynomial normalization of the AIR package; FMG, affine and nonlinear fluid transformation.

Figure 6

Gain in tissue class overlap (in %) provided by FMG as compared to AFF, SPM, and WRP, for each subject, the tissue segmented HBA template serving as a reference. Positive values indicate that FMG provides a better overlap of the tissue type. Negative values indicates that FMG provides a worse overlap. AFF, affine transformation; SPM, SPM linear and nonlinear routine of the SPM package; WRP, affine and fifth order polynomial normalization of the AIR package; FMG, affine and nonlinear fluid transformation.

Figure 7

Gain in misclassification (in %) provided by FMG as compared to AFF, SPM and WRP, for each subject. Negative values indicate that FMG provide a lower percentage of misclassified voxels. Positive values indicates that FMG provides a higher percentage of misclassified voxels. AFF, affine transformation; SPM, SPM linear and nonlinear routine of the SPM package; WRP, affine and fifth order polynomial normalization of the AIR package; FMG, affine and nonlinear fluid transformation.

Figure 8

Impact of spatial normalization procedures on “low resolution” functional activation maps. Stereotactically normalized PET volumes were smoothed with a 3D 8 mm FWHM Gaussian filter. Functional clusters were detected in intersubject averaged (n = 18) PET difference volumes spatially normalized by the four different normalization procedures. Left side: Sagittal and coronal orthogonal maximum intensity projection maps of functional volumes thresholded at a 0.05 significance level corrected for multiple comparisons. Note that the contour delineating the brain is derived from the MNI template whereas the statistical analysis was performed on the HBA template. Right side: Sagittal sections (x = −54, −40, −28, −3, +24 mm) of the functional volumes superimposed on the corresponding HBA template sections, illustrating activations specific or common to the four functional volumes. Color code indicates which functional volume(s) a given voxel belongs to. A, affine transformation; S, SPM linear and nonlinear routine of the SPM package; W, affine and fifth order polynomial normalization of the AIR package; F, affine and nonlinear fluid transformation.

Figure 9

Impact of spatial normalization procedures on “high resolution” functional activation maps. Stereotactically normalized PET volumes were smoothed with a 3D 4 mm FWHM Gaussian filter. Functional clusters were detected in intersubject averaged (n = 18) PET difference volumes spatially normalized by the four different normalization procedures. Left part of the figure shows sagittal and coronal orthogonal maximum intensity projection maps of functional volumes thresholded at a 0.05 significance level (corrected for multiple comparisons). Note that the contour delineating the brain is derived from the MNI template whereas the statistical analysis was performed on the HBA template. Right part shows sagittal sections (x = −54, −40, −28, −3 mm) of functional volumes superimposed on the corresponding HBA template sections, illustrating activations specific or common to the four functional volumes. Color code indicates which functional volume(s) a given voxel belongs to. A, affine transformation; S, SPM linear and nonlinear routine of the SPM package; W, affine and fifth order polynomial normalization of the AIR package; F, affine and nonlinear fluid transformation.