sequence for WMH quantification, a high-resolution 3-D coronal T1-weighted spoiled gradient-echo (SPGR) and a PD/T2-weighted fast spin-echo sequence collected in the same plane as the functional images.
WMH Segmentation and Quantification
Segmentation of WMH volumes was performed on the FLAIR images as described previously (DeCarli, Murphy, Teichberg, Campbell, & Sobering, 1996; DeCarli et al., 1992; Murphy, DeCarli, Schapiro, Rapoport, & Horwitz, 1992). In brief, initial reorientation of the 3-D volume images was performed so that brain regions were accurately delineated using common internal land- marks (Murphy et al., 1993, 1996). Prior to segmenta- tion, nonbrain elements were manually removed from the image by operator-guided tracing of the dura mat- ter within the cranial vault and image intensity non- uniformity correction was applied (DeCarli et al., 1996). Our method of image segmentation rests on the as- sumption that, within a given 2-D image, image pixel intensities for each tissue type (such as cerebral spinal fluid [CSF] and brain matter, or gray matter and white matter) have their own population distribution that dif- fers, but possibly overlaps with that of the other tissue types.
CSF–brain matter segmentation was obtained by math- ematically modeling the pixel intensity distributions from each image using Gaussian normal distributions as previ- ously described (DeCarli et al., 1992). The optimal seg- mentation threshold was defined as the intersection of the CSF modeled distribution with the brain matter modeled distribution (DeCarli et al., 1992). After image segmentation of brain from CSF was performed, the pixel intensity histogram of the brain-only FLAIR image was modeled as a lognormal distribution, and pixel inten- sities three and one-half standard deviations above the mean were considered WMH (DeCarli et al., 1995).
Each subject’s FLAIR and segmented WMH image were then linearly aligned to his or her high-resolution T1 image, and the T1 image was spatially normalized to a minimal deformation target (MDT) (see below for details on spatial normalization and the MDT). Each subject’s T1 to MDT warping parameters were then ap- plied to their segmented WMH image to bring it into MDT space. To measure global WMH volume, total WMH volume was normalized to the MDT volume for each subject. The data were then log transformed be- cause the distribution of WMH volume/brain volume was positively skewed.
The dorsal PFC region was then delineated on the MDT as described previously (Tullberg et al., 2004). In brief, a ray-casting program was used to create different ROIs. The dorsal PFC region was created by casting three rays: (1) one ray along the axis of the anterior and posterior commissure, (2) a second ray parallel to the first, but at the superior boundary of the callosal body,
and (3) a third ray running perpendicular from ray 1 at the point of the anterior commissure. The dorsal PFC region was delineated as the volume resulting from the intersection of rays 2 and 3. The resulting region includ- ed the superior frontal gyrus and the superior portion of the middle frontal gyrus (BA 8 and 9 and the superior portion of BA 10 and 46). Dorsal PFC WMH volumes were calculated from the underlying white matter of this region by counting the number of voxels on each subject’s segmented WMH image that fell within this region. Volumes for left and right hemispheres were added together to determine the regional dorsal PFC WMH volume for each individual.
fMRI Data Preprocessing and Spatial Normalization
Functional imaging data were realigned in SPM99 and spatially normalized using in-house, atlas-based, high-dimensional nonlinear warping procedure (cubic B-splines) and spatially smoothed with an 8-mm full width half maximum Gaussian filter. Due to structural brain changes, such as atrophy, that are characteristic of aging brains (Salat et al., 2004; Good et al., 2001), we did not use the standard MNI template (an average of MRIs from 152 young subjects) as a target for spatial normalization. Instead, we derived an MDT image, an anatomically detailed synthetic image to be used as a target for spatial normalization. By using the MDT as a template, we were able to minimize the total de- formations that result when warping the template onto each subject of that data set. Moreover, the nonlinear warping techniques used here allow for independent adjustment of local matches, resulting in preservation of anatomical detail. Accordingly, this procedure maxi- mized our sensitivity to detect activations in across- subject analyses.
The MDT image was derived as follows: First, an arbitrarily selected image from the study was used as a preliminary target and warped onto each of the subject images. The average deformation of all warps from the target to each subject was computed. Next, the prelim- inary target was deformed by this average deformation to produce the minimal deformation template. The subject images were again normalized, this time to the minimal deformation target.
The warping method was a multigrid application of cubic B-splines. A grid of equally spaced control points enables locally independent warps to be constructed in small subvolumes defined by cubes having control points as vertices. These result in a matching of fine anatomical details. Each data voxel in the target and subject image is contained within a 4 4 4 cube of such control points, and its position is defined by a sum of tensor products of B-spline basis functions (third order polynomials) together with the positions of these control points. The third-order polynomial basis
Nordahl et al.