Distribution of Cortical Benzodiazepine Receptor Binding in Right-Handed Healthy Humans: A Voxel-Based Statistical Analysis of Iodine 123 Iomazenil SPECT with Partial Volume Correction

SPECT

In each subject, 222 MBq of I-123 iomazenil was intravenously administered. SPECT images were acquired twice sequentially to quantify the receptor binding by using a simplified method.⁶ Early SPECT images were acquired immediately after injection, and delayed SPECT images were acquired 162 minutes after injection (the midscan time of the delayed scan was 180 minutes after injection). Venous blood samples were drawn from the arm contralateral to the injection site at 30 minutes postinjection. The SPECT scanning was performed by using a 4-headed gamma camera⁷ (Gamma View SPECT 2000H; Hitachi Medical, Tokyo, Japan) with a low-energy middle-resolution thin-section parallel-hole collimator. SPECT images were acquired every 22.5 seconds until 36 minutes, with 96 collections over 360°, and the data were recorded in a 64 × 64 matrix. The raw SPECT data were added to make an image and transferred to a nuclear medicine computer (HARP 3; Hitachi Medical). The projection data were prefiltered by using a Butterworth filter (cutoff frequency, 0.20 cycles/pixel; order, 10) and reconstructed into transaxial sections of 4.0-mm-thick images in planes parallel to the orbitomeatal line by using the filtered back-projection method. Attenuation correction was performed by using the Chang method,⁸ with an optimized effective attenuation coefficient of 0.12 cm⁻¹.

A quantitative analysis was performed by using the table look-up procedure, based on the 3-compartment 2-parameter model.⁶ We adopted a noninvasive version of this method by using a standardized arterial input function and a single venous blood sample,⁹ to minimize the invasiveness of the procedure. The measurement of the absolute I-123 iomazenil BP in healthy humans is reasonably accurate by using this model, and the method has been validated in previous studies.^6,9

Briefly, parameters K₁/k₂ were fixed at 3.00, and k₄ was fixed at 0.26 throughout the study. The arterial input function was estimated by using the venous blood sample, as described above. On the basis of these fixed parameters and the arterial input function, 2 look-up tables were generated, in which K₁ and k₃ were uniquely determined. As previously described,¹⁰ the I-123 iomazenil BP can be calculated according to the following equation: where f₁ is the free fraction of the parent compound in the plasma, which was fixed at 0.24 as determined in a previous study.⁶

MR Imaging

MR imaging was performed by using a Signa Excite 3T scanner (GE Healthcare, Milwaukee, Wisconsin). A 3D structural MR image was acquired for each subject by using a T1-weighted spoiled gradient-recalled echo sequence (TR/TE, 1.916/8.68 ms; flip angle, 18°; matrix size, 512 × 512; yield, 248 axial sections; section thickness, 1.4 mm; in-plane resolution, 0.43 × 0.43 mm) and T2-weighted 2D fast spin-echo sequences (axial plane; FOV, 250 mm; matrix, 512 × 512; section thickness, 5 mm; intersection gap, 1–1.5 mm; TE, 90–131 ms; TR, 4500–5000 ms).

Regional Gray Matter Concentration Maps

All the procedures were performed by using a personal computer (Dell Dimension 8300; Dell, Round Rock, Texas) running Microsoft Windows XP (Microsoft, Redmond, Washington).

The 3D T1-weighted MR image was resliced in the native space of each subject by using a voxel size of 1.0 × 1.0 × 1.0 mm. The results were first segmented into gray matter, white matter, and CSF and were then spatially normalized by using the unified model¹¹ of SPM5 (Wellcome Department of Imaging Neuroscience: http://www.fil.ion.ucl.ac.uk/spm/) according to the optimized VBM protocol.^3,12 In this study, however, we skipped the ‘‘modulation” step by multiplying the gray matter maps by the Jacobian determinants of the corresponding spatial normalization matrices, which corrects the local volumetric contraction or dilation that the images undergo during warping. These procedures generated both spatial normalization matrices and inverse spatial normalization matrices. The normalized map of the gray matter yielded in the above procedures was transformed into native space by using an inverse spatial normalization matrix. The resulting gray matter map in the native space was convoluted with a 3D Gaussian function with a FWHM of 12 × 12 × 12 mm for the purpose of SPECT PVE correction.^13,14 The FWHM was assumed to be the same as the PSF of the reconstructed SPECT image, which was assessed by using an I-123 one-millimeter-diameter line source in air, according to a previously described methodology.⁷ The resultant image was subsequently referred to as the “rGMC map,” which was also used in the SPECT PVE correction process as described in previous studies.^13,14

In this study, the smoothing process for the gray matter maps obtained by using MR imaging was performed not in the standard space, but in the native space. The rationale for this step was that if the normalization process was performed perfectly, then the individual differences in the thickness of the cortical ribbon on the MR imaging findings in the native space would be cancelled by those in the standard space. For the purpose of voxel-based statistical analysis, the rGMC map was spatially normalized by the spatial normalization matrices generated above. The modulation step, as mentioned above, was skipped to preserve the voxel value of the rGMC map after spatial normalization.

PVE Correction

The PVE of the acquired SPECT images influences the estimated parameters. In a 3-compartmental model, the tissue activity B(t) is given by the formula: where C_a(t) is the arterial input function and f(t) is the transfer function given by the formula: where

The PVE-uncorrected tissue activity observed at a certain voxel B′(t) is approximately indicated as follows²: where g is the rGMC value at the voxel. Because g is a time-independent factor, according to the equations 2, 3, and 4, the parameter K′₁ corresponding to B′(t) is given by the formula: whereas other parameters k´_t(i-2, 3, 4) corresponding to B′(t) are identical with k_i. According to equation 1, BP′ is given by the equation:

The equations 5 and 7 demonstrate that the PVE for BP images can be expressed approximately in the same fashion as that for the acquired SPECT count images (Fig 1).² PVE correction for the BP images, therefore, was performed by using 3D T1-weighted MR imaging data by using a method that was basically the same as our previously presented method for qualitative I-123 iomazenil SPECT images (Fig 1).² The I-123 iomazenil BP images were coregistered with the rGMC maps by using the Linear Image Registration Tool (FLIRT) of Functional MR Imaging of the Brain (http://www.fmrib.ox.ac.uk/).¹⁵ In this procedure, the BP images were simultaneously reformatted to matrices with the same voxel size as the referenced rGMC maps in the native space. The precision of coregisteration was inspected by using the “Check Registration” tool in SPM5.

• Download figure
• Open in new tab
• Download powerpoint

Fig 1.

A, I-123 iomazenil SPECT image. B, Automatic coregistration of the I-123 SPECT image with the MR image via smoothed gray matter maps. The maps are simultaneously reformatted to a matrix that is the same size as the referenced smoothed gray matter map. C, 3D MR image obtained before surgery. D, MR image segmented into Bayesian probability maps showing 3 tissue classes (gray matter, white matter, and CSF maps). E, The gray matter map convoluted with the PSF, which is assumed to be the same as the PSF of the SPECT scanner. The resultant image is subsequently referred to as the rGMC map. F, Smoothed gray matter map masked with a threshold set to 35% of the maximum voxel value. The coregistered I-123 SPECT image is divided by using the masked smoothed gray matter map on a voxel-by-voxel basis. G, Image (B) anatomically normalized by the spatial normalization matrices generated in the segmentation process. H, Image (F) anatomically normalized in the same manner as in image G.

Binary mask images for the gray matter were created based on the rGMC maps to eliminate voxels with a very low rGMC (ie, voxels located at irrelevant positions away from the gray matter), because such voxels might amplify the noise in the PVE correction. The threshold for determining the boundary of the binary mask images was set even lower (1% of the maximum rGMC, an empirically determined value) than that used in previous studies^2,13 to minimize the error produced by the smoothing of the masked images before the voxel-based statistical estimation. These masked images based on the rGMC were then applied to the coregistered BP images. The masked BP images were then divided by using the rGMC map on a pixel-by-pixel basis. The threshold for masking, on the other hand, was set at 35% of the maximum rGMC (the same as the threshold value used in previous studies^2,13) when displaying the PVE-corrected BP images (Fig 1).

Automated VOI Analysis

Before the VOI analysis, each SPECT image was coregistered with the corresponding rGMC map and then spatially normalized by using the same normalization matrices as those generated in the above-mentioned normalization process for the rGMC map. Automated VOI analysis was performed for the normalized images by using an in-house tool based on normalized VOI templates. With this tool, voxels with a zero or negative value were automatically eliminated from the VOI to avoid selection of irrelevant voxels. In the present study, the AAL atlas¹⁶ was used as the template for the VOI analysis (Fig 2). Automated VOI analyses for the rGMC map and PVE-uncorrected BP maps were performed to elucidate the effect of PVE associated with age-related rGMC changes on the PVE-uncorrected SPECT BP images (Fig 3). The subjects were divided into 2 groups: the young group (40 years of age or younger) and the elderly group (older than 60 years). The correlations between the young-to-elderly ratio of the group-averaged rGMC and that of the group-averaged PVE-uncorrected BP were assessed for all 116 AAL regions.

• Download figure
• Open in new tab
• Download powerpoint

Fig 2.

Anatomically normalized VOI set extracted from the VOI template based on AAL. See the text for details.

• Download figure
• Open in new tab
• Download powerpoint

Fig 3.

Flow diagram showing the preprocessing steps for the VOI analysis and voxel-based statistical analysis protocols.

Voxel-Based Statistical Analysis

Before the voxel-based statistical analysis, all the SPECT images were spatially normalized in the same way as mentioned above (Fig 3). All the normalized images prepared for the SPM analysis were smoothed by using a 12-mm FWHM isotropic Gaussian kernel. This step conditions the residuals to conform more closely to the Gaussian random field model underlying the statistical process used for adjusting the P values.¹⁷ After smoothing, the normalized binary mask images were also applied to the analysis as explicit mask images, to exclude irrelevant voxels corresponding to an rGMC of <35% of the maximum from the analysis. Global scaling was not applied to the voxel-based statistical analysis used in the present study.

The SPM statistical model used voxel-by-voxel multiple regression analyses to detect voxels showing a significant correlation between the rGMC and age (ie, VBM) or SPECT BP and age. In these models, age was included as a covariate of interest and sex was cast as a nuisance variable. The resultant set of voxel values constituted a statistical parametric map of the t statistic t-score map. The significance threshold was set at P < .05, with family-wise error correction for the peak height; no cluster size threshold was set.