Magnetic resonance images were acquired using a scanner (Philips Gyroscan; Philips Medical Systems, Best, the Netherlands) operating at 1.5 T in all subjects. T1-weighted, 3-dimensional, fast field echo scans with 160 to 180 1.2-mm contiguous coronal slices (echo time [TE], 4.6 milliseconds; repetition time, 30 milliseconds; flip angle, 30°; field of view, 256 mm; and in-plane voxel sizes, 1 × 1 mm2) and T2-weighted, dual echo turbo spin echo scans with 120 1.6-mm contiguous coronal slices (TE1, 14 milliseconds; TE2, 80 milliseconds; repetition time, 6350 milliseconds; flip angle, 90°; field of view, 256 mm; and in-plane voxel sizes, 1 × 1 mm2) of the whole head were used for quantitative measurements. In addition, T2-weighted, dual echo turbo spin echo scans with 17 axial 5-mm slices and a 1.2-mm gap (TE1, 9 milliseconds; TE2, 100 milliseconds; flip angle, 90°; field of view, 250 mm; and in-plane voxel sizes, 0.98 × 0.98 mm2) were used for clinical neurodiagnostic evaluation. Processing was done on the neuroimaging computer network of the Department of Psychiatry, which includes workstations (Unix 9000; Hewlett Packard, Palo Alto, Calif), a computer server, and Pentium III–equipped personal computers. Prior to quantitative assessments 10 MRIs were randomly chosen and cloned for intrarater reliability determined by the intraclass correlation coefficient. All MRIs were coded to ensure masking for subject identification and diagnosis, scans were put into a Talairach frame (no scaling), and corrected for inhomogeneities in the magnetic field.27 Binary masks of gray matter were made based on histogram analyses and a series of mathematical morphological operators to connect all voxels of interest within the cranium, as validated previously.28 The binary gray matter masks were then analyzed using voxel-based morphometry. The binary gray matter masks were resampled to a voxel size of 2 × 2 × 2.4 mm3, blurred using an isotropic Gaussian kernel (full width at half maximum of 8 mm) to generate gray matter "density maps." The density maps represent the local concentration of gray matter (between 0 and 1) per voxel. Each of the MRIs was transformed into a standardized coordinate system in a 2-stage process using the ANIMAL algorithm.29 In the first step, a linear transformation was found by minimizing a mutual information joint entropy objective function computed on the gray level images.30 A nonlinear transformation was computed in the second step by maximizing the correlation of the subject's image with that of a standardized brain. The nonlinear transformation is run up to a scale (full width at half maximum of 4 mm) that aligns global anatomical regions while minimally affecting local volume changes. The standardized brain was selected earlier among 200 brain MRIs of healthy subjects between the ages of 16 and 70 years. To select the standardized brain, all 200 brain MRIs were registered to the Montreal standard brain31 and averaged, yielding one average brain image. The mean square error on the normalized intensity values was computed between each of the brain MRIs and the average brain image. The standardized brain was the brain image with the smallest mean square error. Transformations were then applied to the gray matter density maps to remove global differences in the size and shape of individual brains.