点击显示 收起
the Department of Neurology (C.D., E.F., V.R., W.J.J.) and Imaging of Dementia and Aging (IDeA) Laboratory (C.D., E.F., V.R., W.J.J.), Center for Neuroscience, and the Division of Biostatistics (D.H.), Department of Epidemiology and Preventive Medicine, University of California at Davis, Davis, Calif.
Abstract
Background and Purpose— MRI segmentation and mapping techniques were used to assess evidence in support of categorical distinctions between periventricular white matter hyperintensities (PVWMH) and deep WMH (DWMH). Qualitative MRI studies generally identify 2 categories of WMH on the basis of anatomical localization. Separate pathophysiologies and behavioral consequences are often attributed to these 2 classes of WMH. However, evidence to support these empirical distinctions has not been rigorously sought.
Methods— MRI analysis of 55 subjects included quantification of WMH volume, mapping onto a common anatomical image, and spatial localization of each WMH voxel. WMH locations were then divided into PVWMH and DWMH on the basis of distance from the lateral ventricles and correlations, with total WMH volume determined. Periventricular distance histograms of WMH voxels were also calculated.
Results— PVWMH and DWMH were highly correlated with total WMH (R2>0.95) and with each other (R2>0.87). Mapping of all WMH revealed smooth expansion from around central cerebrospinal fluid spaces into more distal cerebral white matter with increasing WMH volume.
Conclusion— PVWMH, DWMH, and total WMH are highly correlated with each other. Moreover, spatial analysis failed to identify distinct subpopulations for PVWMH and DWMH. These results suggest that categorical distinctions between PVWMH and DWMH may be arbitrary, and conclusions regarding individual relationships between causal factors or behavior for PVWMH and DWMH may more accurately reflect total WMH volume relationships.
Key Words: cerebrovascular disorders magnetic resonanace imaging white matter
Introduction
White matter hyperintensities (WMH) are commonly seen on T2-weighted MRI and are often divided into 2 categories: periventricular WMH (PVWMH), which abut the cerebral ventricles, and deep WMH (DWMH), which are patchy areas of WMH in subcortical white matter distinct from the periventricular area.1–8 Qualitative MRI studies evaluating the impact of vascular risk factors on WMH routinely distinguish PVWMH from DWMH.6,9–14 Results from these studies generally show age and vascular risk factors as the strongest correlate of PVWMH, whereas associations between vascular risk factors and DWMH are much weaker. Similarly, studies examining the relationship between PVWMH, DWMH, and cognitive performance among nondemented elderly15 generally find strong correlations between PVWMH and cognitive measures15,16 but not DWMH. MRI pathological correlations of WMH also suggest differences between PVWMH and DWMH.9,17–26 However, within both types of WMH lesions, there is vascular fibrosis and lipohyalinosis,9,22,23,25–27 supporting a common ischemic vascular pathological mechanism for WMH among older individuals.17,24,25 Therefore, whereas qualitative MRI studies generally support distinctions between PVWMH and DWMH, pathological studies suggest that both types of WMH share the same ischemic etiology supporting pathological linkage. However, most previous MRI work has used qualitative single-slice assessments that may not fully appreciate the complex 3D anatomy of WMH. Thus, existing MRI data cannot unequivocally support distinctions between PVWMH and DWMH. This study sought to confirm these anatomical distinctions using new image segmentation and mapping techniques.
Methods
Subjects
Subjects for this study consisted of the first 55 consecutive individuals recruited through the University of California at Davis (UCD) Alzheimer’s Disease Center for whom research MRI was available for analysis. As expected, these individuals had variable cognitive abilities ranging from normal to cognitive impairment not demented (CIND) to dementia as defined according to standard diagnostic criteria.28,29 Etiologies of cognitive impairment included Alzheimer’s disease (AD) and cerebrovascular disease (CVD), including symptomatic stroke, although individuals with cortical infarctions were excluded. Subjects were recruited for participation through advertisements, community screening, and physician referrals. Subject demographics according to cognitive syndrome are summarized in the table. Informed consent was obtained for each patient at the time of participation in the study according to UCD institutional review board guidelines.
Subject Demographics, WMH, and Brain Volumes
MRI Sequences
All brain imaging was obtained at the UCD MRI research center on a 1.5T GE Signa Horizon LX Echospeed system. Two sequences were used: a T1-weighted coronal 3D spoiled gradient recalled echo acquisition and a fluid-attenuated inversion recovery (FLAIR) sequence designed to enhance WMH segmentation.30
Image Analysis
An overview of image analysis is summarized in Figure 1. In brief, image segmentation using previously described algorithms31,32 was applied to the FLAIR sequences to segment WMH (Figure 2). After affine coregistration of the FLAIR image to the high-resolution T1 image, WMH voxels were used to correct intensity changes in the T1 image to reduce any adverse impact of the WMH voxel values on the accuracy of the nonlinear warping algorithm. The details and rationale for these processes are included in an appendix available online only at http://www.strokeaha.org.
Data Analysis
Nonlinear warping enables precise matching of anatomical regions across subjects (see online appendix). We used this characteristic of the method to determine the exact distance between each WMH voxel and the ventricular ependymal surface for all subjects. To test the hypothesis of the PVWMH versus DWMH distinction, we measured distributions of WMH voxels in reference to the ependymal surface of the target ventricular system in 2 ways. We first created histograms of the average distance from the ventricular surface for 5 quintiles of WMH burden. We hypothesized that if a true distinction in WMH location (ie, PVWMH versus DWMH) were present, we would see 2 peaks in the histograms related to the separate WMH categories. Second, we created a standardized division of WMH location into PVWMH and DWMH on the basis of 1-cm distance from the ventricular system. Volumes for PVWMH and DWMH were then calculated for each individual subject. Linear regression analysis was used to examine PVWMH and DWMH volumes in relation to total WMH volume and each other.
Results
Subjects
There were no significant differences in age across the groups, although subjects with dementia tended to be somewhat older (Table). Among the dementia subjects, 9 were diagnosed as clinically probable AD, and 6 were diagnosed as mixed dementia with AD and CVD combined. A total of 8 subjects had clinical stroke that presented as a lacunar syndrome, predominantly with hemiparesis. Of the 8 with clinical stroke, 6 were demented and 2 had CIND. No subjects had cerebral hemorrhage.
WMH Volumes
WMH volumes were calculated for all subjects and ranged from 1.1 to 63 mL and divided into quintiles with ranges consisting of first, 1.1 to 2.3 mL; second, 2.6 to 4.6 mL; third, 5.1 to 9.0 mL; fourth, 9.7 to 16.0 mL; and fifth, 18.0 to 63 mL. There were no significant differences in mean WMH volumes in association with degree of cognitive impairment (Table), although subjects with dementia had nearly twice the volume of WMH.
WMH Distributions
Mapping of subjects by quintile of total WMH volume (Figure 3) revealed a continuous gradient of mapped WMH voxels extending from around the cerebrospinal fluid (CSF) ventricular system in direct relation to calculated WMH volume. Evidence for the potential misclassification of DWMH on the basis of 2D visualization is also illustrated (Figure 3). When viewed axially, as is common in studies of WMH,33,34 DWMH appear present. However, the sagital and coronal orientations show that these WMH are actually contiguous with the ventricular lining.
Distance histograms of WMH voxels are shown in Figure 4. There is no clear sign of a bimodal distribution. Instead, the peak of the WMH distribution widens continuously from the lowest WMH quintile, where the median distance is 3.5 mm, to highest quintile, where the median distance is 6.0 mm. One exception to this general observation is at the lowest quintile, where a small second peak occurs at 30 mm from the ventricular surface. Examination of the images in the lowest quartile of WMH revealed the presence of multiple punctate WMH scattered within the centrum semiovale.
In the second analysis, WMH were divided into PVWMH and DWMH on the basis of a 1-cm distance from the ventricular surface. The relationship between PVWMH, DWMH, and total WMH burden is graphically illustrated in Figure 5. PVWMH and DWMH volumes were closely associated with WMH burden (R2=0.99 and 0.92, respectively). PVWMH and DWMH volumes also were significantly correlated (R2=0.87). The slope of PVWMH to total WMH burden is 2.5x that of the slope between DWMH and total WMH burden, suggesting a preferential increase in PVWMH with increasing total WMH burden.
Discussion
Use of image segmentation, 3D anatomical mapping of WMH voxels, and 2 separate analytical methods failed to find distinctions between PVWMH and DWMH. Not only did analyses of distance histograms fail to identify 2 separate WMH voxel populations, but application of a standard categorical definition for PVWMH versus DWMH across all subjects found high correlations with total WMH burden as well as with each other. These results suggest that categorical distinctions between PVWMH and DWMH are likely to be arbitrary, and conclusions regarding individual relationships between causal factors or behavior for PVWMH and DWMH may more accurately reflect total WMH volume relationships. However, our data cannot speak to possible regional differences in pathological processes for PVWMH and DWMH but do show that both phenomena are highly correlated with each other, suggesting a common underlying mechanism.
Our results appear different from visual inspection (Figure 2) as well as published examples of WMH.8,33 One obvious explanation for this discrepancy is our use of 3D mapping techniques that may avoid some of the limitations of 2D qualitative MRI studies. For example, we show that WMH typical of DWMH, when viewed axially, are in fact contiguous with ventricular WMH (Figure 3). This finding is not specific to our method because a recently published Statistical Parametric Mapping study found similar results,35 although it did not specifically examine the question of PVWMH versus DWMH. A second explanation for differing results also may derive from our use of consistent measures and anatomical definition of PVWMH versus DWMH. For example, in the Rotterdam Scan Study, DWMH are measured according to width and number as opposed to categorical definitions for PVWMH,8 making direct comparisons between the 2 types of WMH difficult. Different measures for DWMH versus PVWMH may also explain differences in associations between causal factors, behavior, and DWMH found with qualitative studies.15,16 However, our conclusions are not meant to suggest that islands of abnormal WMH signal located in the centrum semiovale do not exist. Quite the contrary; we believe that our data support the notion proposed by Schmidt et al36 that WMH burden increases through the confluence of PVWMH with punctate WMH located in the centrum semiovale, although our experiment was not designed to address this particular question. However, our histogram data do support this hypothesis by showing a second peak at the lowest quintile of WMH volume, indicating more frequent DWMH initially that may then converge with PVWH as the total WMH increases. We further suggest that the strong correlations between causal factors and behavior found with PVWMH in qualitative studies likely reflect the steeper slope of change of PVWMH with total WMH burden, as seen with our quantitative data analysis (ie, the steeper slope suggests increased sensitivity to detect differences across individuals).
Our results are also consistent with current concepts of WMH pathology. Although some controversy remains,37 there is general consensus for a single vascular white matter watershed area extending between 3 and 13 mm from the ventricular surface,37–40 remarkably similar to the distances described by our quantitative MRI analysis (Figure 4). However, some neuropathological evidence distinguishing different types of WMH lesions does remain.25 For example, subependymal gliosis, irregularity of the ependymal lining, adjacent myelin pallor,22,23,25,41,42 or a normal fasciculus subcallosus22 are commonly found in postmortem samples when WMH are limited to ventricular capping or a smooth halo about the ventricles (eg, similar to WMH quintiles 1 through 3; Figure 3). Conversely, vascular hyalinization, ischemic white matter injury, and microscopic infarction are consistently found when the periventricular changes become extensive.17,24,25 In these cases, DWMH sharing features of ischemic pathology commonly co-occur with PVWMH,25 suggesting a pathophysiology common to both.17,24 Although differences in pathological features argue strongly for separate categories of WMH, we believe these categories are different from designations of PVWMH or DWMH used for qualitative MRI studies. That is, minor degrees of WMH (rims and caps1) are most consistent with periventricular edema or disturbed CSF transport25,41,42 and most likely accompany normal aging.32 Conversely, more extensive WMH likely have a vascular etiology independent of designations such as PVWMH or DWMH.17,24,25 Quantitative MRI studies support this distinction by showing strong associations between vascular risk factors and vascular disease when WMH volumes are extensive,32,43 further supporting the notion that it is the overall extent and not categorical distinctions that best represent the underlying pathology of the WMH.
Because of a number of limiting factors, these results should be interpreted cautiously. First, our data are based on segmented WMH values and mathematical interpolation methods, raising the possibility that measurement error might reduce the sensitivity to identify a second population of WMH voxels. That is, our segmentation method favors selecting voxels of most extreme signal change, and our interpolation method induced a small amount of image smoothing. However, we do not believe that these errors were substantial because our segmentation of WMH is based strictly on voxel intensity parameters31,43 and, therefore, would tend to underestimate the continuity of these changes by selecting only voxels above a specified threshold, favoring separate populations of WMH voxels. Minor degrees of image smoothing from image interpolation during warping are similarly unlikely to lead to substantial error in detection of separate voxel populations because the distinctions between PVWMH and DWMH are defined in more macroscopic terms.15,34 Subject selection may be a second weakness of the study for 2 reasons. First, our population included individuals referred to or recruited for our memory disorders clinic and, therefore, is not representative of the general population. However, the patterns of WMH seen with the 55 subjects studied do not differ from other studies reporting various degrees of WMH severity.33,34 Secondly, our sample included individuals with a wide range of cognitive abilities and concurrent CVD, although individuals with cortical infarction were excluded from the analysis. Although it has been suggested that individuals with cognitive impairment are more likely to have larger PVWMH compared with DWMH,44 we would argue that this finding more closely reflects total WMH burden, as seen with other studies.45,46 Therefore, although this is a sample of convenience that included individuals with differing degrees of cognitive impairment, the type and pattern of WMH seen were typical of those noted by population studies and would not be expected to alter the results found. However, our conclusions may not be directly applicable to other diseases such as late-life depression, in which frontal DWMH are significantly more common,47,48 an area worthy of further investigation using these newer methods.
In conclusion, we believe the methods developed here conclusively show that WMH extend smoothly from the ventricular wall as the overall burden increases, offering no clear evidence for distinguishing WMH subtypes. In fact, these data support the notion of a single vascular watershed area that extends from the CSF ventricular surface to the central white matter, consistent with currently proposed cerebral vascular anatomy.37,38 These observations also support the notion of a common ischemic etiology among elderly individuals when WMH burden is extensive.
Appendix
Method of Nonlinear Warping
General Methods
WMH Segmentation
Image evaluation was based on a semiautomatic segmentation analysis that involves operator-guided removal of nonbrain elements as described previously31,49 and illustrated43 with the added modification of morphometric erosion of 2 image pixels before modeling to remove the effects of partial volume CSF pixels on WMH determination.
Affine Transformation of Images
Affine transformation of the FLAIR image to each subject’s T1 image and initial registration of subject T1 images to the target image used a 12-parameter method.50 This method was modified for use of normalized mutual information51 proven to be robust when matching images of different modalities.
T1 Image Modification for WMH Signal Effects
This multistep process involved applying the affine transformation of the individual’s FLAIR image to the T1 image and using the coregistered and segmented WMH map to identify voxels of low signal intensity in the white matter of the T1 image for intensity replacement. Thresholding at 128-pixel intensity values (one half the original 255) of the WMH map after transformation from the FLAIR image to the T1 image was used to reduce interpolation error. Replacement pixel intensities were estimated from the mean voxel intensities for nonlesioned white matter immediately surrounding the identified WMH after adding Gaussian noise to the intensity values.
Imaging Mapping
Nonlinear transformation was applied to the images after skull removal and affine alignment of the subject to the target image. We followed the method described by Otte,52 with the modifications described below.
Spline Matching
Overview of Trilinear Spline Matching
Spline matching of a subject image to a target image uses initially identical control grid lattices placed over each image. Let P denote the control lattice overlaying the subject image and Q its companion over the target image. P and Q are initially evenly spaced lattices. P is not changed, but points of Q will be adjusted during optimization. A voxel location of the underlying image can be expressed by barycentric coordinates in terms of the 8-corner control lattice points of the lattice cube in which it lies. Each point in P or Q is the corner of 8 cubical subvolumes of the underlying image defined by points in the control lattice. When a point is moved in Q, it affects the shape of the 8 subvolumes to which it is a corner. This creates a local nonlinear warping by changing the positions of image voxel locations lying within the corresponding cubical volumes in P.
The Iterative Conditional Modes (ICM)52,53 approach was used to optimize a match between images as described below. In this method, each of the Q control lattice points is tested for improving the image match by moving it toward each of its neighbor control points in all 6 directions. The movement (or lack of movement) creating the B distortion with the best local match is retained, and the algorithm then repeats on another control point, eventually testing them all, in random order.
Optimization of image matching used a multiresolution and multigrid approach that used a hierarchy of decreasingly smoothed images using progressively smaller control grid spacing.
Local Patch Warping With 3D Trilinear Bezier Splines
Assume a control grid lattice P is placed over the object image, and another initially identical lattice Q is over the target image as described previously. To simplify the notation, assume that the distance between each of the P points is 1. Now let V=(x, y, z) be the location of a data voxel in the image to be warped. We can find the corner points of the cube in which V lies by locating the local "origin" p000=(x, y, z), where designates the greatest integer less than or equal to its argument. The 8 points of the cube containing V will then be the set of lattice points {pijk, i=0,1, j=0,1, k=0,1}, where nonzero values of i, j, k indicate adding 1 to the corresponding coordinate of p000.
To obtain the barycentric coordinates of V with respect to its 8 control lattice points, define the first-order spline basis functions in Equation 1, where 0t1. (1) B0(t):=1–t and B1(t):=t.
Let D=(s, t, u)=V–p000. This is the vector of offsets of V from the local origin p000. Each component of D is bounded between 0 and 1 (in the general case when the lattice spacing of P differs from 1, the offsets are adjusted to be fractional distances in the coordinate directions from p000 to its neighbors). The barycentric coordinates of V with respect to its bounding cube corners can be expressed as the 3D tensor product of the basis functions using D as described in Equation 2. (2) V=i=0,1j=0,1k=0,1pijkBi(s)Bj(t)Bk(u).
The set of all image points for these pijk and values of s, t, u between 0 and 1 constitutes a local 3D volume Bezier spline control patch. Local image transformation occurs when points in Q corresponding to 1 of the pijk are moved. That is, replacing each pijk by its relocated companion qijk in Equation 2 changes the shape of the local control patch and moves points with the same barycentric coordinates to new physical locations. The location of V in the target image is thus transformed to a new position, the world coordinates of which can be computed from the locations of the qijk.
The spline transformation of the image is the sum of the transformations of each of its control patches. It is continuous across the boundaries of the patches because boundary locations have a value of 0 or 1 in 1 of the offset components s, t, or u. A boundary location having a 0 (1) offset in 1 patch is equivalently expressed with an offset of 1 (0) in the appropriate adjoining patch, and these offsets correspond to the shared control points at the boundary of the patches.
Optimization Algorithm
Spline matching between 2 images involves moving the Q control lattice points to maximize the match between target and object. As alluded to above, optimization was based on the ICM algorithm introduced by Besag53 as implemented for trilinear spline transformations by Otte.52 With this approach, a randomly chosen point in Q is individually moved one quarter of the distance to each of its neighbors, and the move (or lack of movement) creating the best improvement in image matching is retained as the new location of that control point. This is repeated for each of the Q points, the set of trials for all points constituting 1 iteration. When all the points of Q have been assessed in this manner, then a subsequent iteration is started using a different random order of Q points to minimize bias. Iterations continue until the cumulative improvement in the match falls below a threshold value.
Multigrid and Multiresolution Approach
To achieve the best matching of large and small features, we used a multigrid and multiresolution approach.51 Using multiple grids involved a sequence of optimizations using P and Q control grids of decreasing lattice point spatial separation. Multiple resolution involved used of a Gaussian pyramid52 of smoothed images in conjunction with our hierarchy of grid spacings.
For this study, the multigrid optimization involved a sequence of P grid lattice distances starting at 32-mm separation, then 25, 15, 10, 5, 3, and 2 mm. After the initial matching, the Q grid was transformed using the optimal matching parameters computed at the previous coarser grid spacing.
The multiresolution process began with oversampled images having isotropic voxels of 0.35 mm on a side (the "native" voxel size was typically 0.9766x0.9766x1.5 mm). Oversampling of the T1 images was necessary to provide enough voxel location points for accurate volume matching using the finest-mesh control grids, as described below.
Image smoothing for matching with the 32- and 25-mm grids involved 2 convolutions: first with a Gaussian kernel isotropic at full width at one half maximum of 1.5 mm and then again at full width at one half maximum of 3 mm. Downsampling of the smoothed image was used to recover isotropic voxels of 1.5 mm on a side. At the 15 and 10 mm grid sizes, we smoothed the oversampled image once using isotropic full width at half maximum of 1.5 mm and then undersampled to make an image with isotropic voxels of this size. Finally, at 5-, 3-, and 2-mm grids, we used the unsmoothed, oversampled images (istropic voxels at 0.35 mm) in or to provide sufficient data for our matching criterion at the smallest grid resolution sizes. This approach was based on experiments using optimization on native images at grid sizes of 5 mm that resulted in poor matching because of the small sample sizes for local transformations. For example, with 2-mm control point separation, a cube of control points contains on the average 180 image voxels of size 0.35 mm on a side. This image resolution allows for an estimated total of 1440 voxels (in the 8 local cubes neighboring a control point) at which the matching can be evaluated with movement of each control point. We found this sufficient sampling to obtain a meaningful measure of the local match using our chosen matching criteria at the smaller grid sizes. At the nonoversampled voxel size of 1.5 mm on a side, we would have >64x fewer locations, resulting in statistically unreliable matching scores.
Matching Criterion
To evaluate the local matching created by a movement of control points, we compared the intensity matching of subject and target images at transformed locations of subject image voxels under the local spline transformation. Each of these locations carried its intensity from the subject image, creating an array s of subject voxel intensities. To compare them with the target image, we estimated the target image intensities at the same transformed locations (which usually did not coincide with the location of a target voxel) by trilinear interpolation of the target intensities. This generated an array t of the target image intensities at the transformed locations. These intensity arrays s and t were inputs to the matching criterion. For the purpose of this experiment, we used the intensity correlation coefficient (CC) method. This is defined as follows. Let us assume s and t denote N-dimensional arrays of image intensities as described above. Let μ(s) and μ(t) denote the means of the respective arrays. As denoted by Equation 3, CC is the correlation coefficient between the random variables represented by the vectors s and t. (3) CC=i=1,N[si–μ(s)]x[ti–μ(t)/(s–μ(s) x t–μ(t)]
Alternatively, CC can be viewed as the dot product of 2 vectors, s and t each shifted by its mean and divided by the norms of these 2 vectors. In other words, it is the cosine of the angle (in N-dimensional space) between the 2 vectors, where –1CC1. Using this method, a CC value close to 0 shows little relationship between the images, whereas a CC value close to 1 shows a close matching.
Results
The accuracy of the image mapping technique was assessed in 4 ways. First, segmented brain images from a subset of 23 of the 55 subjects were mapped onto the target. This created an overlap image for which the image intensity reflected the percentage of subject overlap at each voxel (brightest voxels having greatest overlap; Figure I). Second, we calculated and displayed the distribution of overlapping voxels (Figure II). Third, we created 2 separate binary images, displaying 100% and 75% overlap, respectively (Figure II, top). Finally, individual markings of white matter tracts for the anterior commissure, corpus callosum, and internal and external capsule from 4 subjects were mapped onto the target image to verify accuracy of important internal white matter tracts (Figure III).
In Figure I, we see that the majority of the voxels are at the highest intensity (100% overlap). Histogram analysis supports this finding by showing that >52% of subject image pixels were mapped with 100% overlap onto the target (Figure II). In addition, >75% of subject image pixels were also mapped with >75% overlap onto the target. Conversely, <10% of subject image pixels were mapped with <25% overlap. Thresholding of the frequency maps show that the greatest degree of variability occurred in the cortex (Figure II, top), whereas cerebral white matter and the ventricular system, areas of greatest importance to this study, were mapped with nearly 100% overlap. Figure III shows that internal white matter tracts were also accurately mapped to the target.
Acknowledgments
This research was supported by National Institute of Aging grants P30 AG10129 and R01 AG021028.
References
Kertesz A, Black SE, Tokar G, Benke T, Carr T, Nicholson L. Periventricular and subcortical hyperintensities on magnetic resonance imaging. "Rims, caps, and unidentified bright objects." Arch Neurol. 1988; 45: 404–408.
Mirsen TR, Lee DH, Wong CJ, Diaz JF, Fox AJ, Hachinski VC, Merskey H. Clinical correlates of white-matter changes on magnetic resonance imaging scans of the brain. Arch Neurol. 1991; 48: 1015–1021.
Bowen BC, Barker WW, Loewenstein DA, Sheldon J, Duara R. MR signal abnormalities in memory disorder and dementia. AJNR Am J Neuroradiol. 1990; 11: 283–290.
Fazekas F. Magnetic resonance signal abnormaliaties in asymptomatic individuals: their incidence and functional correlates. Eur Neurol. 1989; 29: 164–168.
Lindgren A, Roijer A, Rudling O, Norrving B, Larsson EM, Eskilsson J, Wallin L, Olsson B, Johansson BB. Cerebral lesions on magnetic resonance imaging, heart disease, and vascular risk factors in subjects without stroke. A population-based study. Stroke. 1994; 25: 929–934.
Ylikoski A, Erkinjuntti T, Raininko R, Sarna S, Sulkava R, Tilvis R. White matter hyperintensities on MRI in the neurologically nondiseased elderly. Analysis of cohorts of consecutive subjects aged 55 to 85 years living at home. Stroke. 1995; 26: 1171–1177.
Schmidt R, Fazekas F, Offenbacher H, Dusek T, Zach E, Reinhart B, Grieshofer P, Freidl W, Eber B, Schumacher M. Neuropsychologic correlates of MRI white matter hyperintensities: a study of 150 normal volunteers. Neurology. 1993; 43: 2490–2494.
de Groot JC, de Leeuw FE, Oudkerk M, van Gijn J, Hofman A, Jolles J, Breteler MM. Cerebral white matter lesions and cognitive function: the Rotterdam Scan Study. Ann Neurology. 2000; 47: 145–151.
van Swieten JC, van den Hout JH, van Ketel BA, Hijdra A, Wokke JH. Periventricular lesions in the white matter on magnetic resonance imaging in the elderly. A morphometric correlation with arteriolosclerosis and dilated perivascular spaces. Brain. 1991; 114: 761–764.
Vermeer SE, van Dijk EJ, Koudstaal PJ, Oudkerk M, Hofman A, Clarke R, Breteler MM. Homocysteine, silent brain infarcts, and white matter lesions: the Rotterdam Scan Study. Ann Neurol. 2002; 51: 285–289.
de Leeuw FE, De Groot JC, Oudkerk M, Witteman JC, Hofman A, van Gijn J, Breteler MM. Aortic atherosclerosis at middle age predicts cerebral white matter lesions in the elderly. Stroke. 2000; 31: 425–429.
de Leeuw FE, de Groot JC, Oudkerk M, Kors JA, Hofman A, van Gijn J, Breteler MM. Atrial fibrillation and the risk of cerebral white matter lesions. Neurology. 2000; 54: 1795–1801.
de Leeuw FE, de Groot JC, Oudkerk M, Witteman JC, Hofman A, van Gijn J, Breteler MM. Hypertension and cerebral white matter lesions in a prospective cohort study. Brain. 2002; 125: 765–772.
Vermeer SE, Hollander M, van Dijk EJ, Hofman A, Koudstaal PJ, Breteler MM; Rotterdam Scan Study. Silent brain infarcts and white matter lesions increase stroke risk in the general population: the Rotterdam Scan Study. Stroke. 2003; 34: 1126–1129.
de Groot JC, de Leeuw FE, Oudkerk M, Hofman A, Jolles J, Breteler MM. Cerebral white matter lesions and subjective cognitive dysfunction: the Rotterdam Scan Study. Neurology. 2001; 56: 1539–1545.
Ylikoski R, Ylikoski A, Erkinjuntti T, Sulkava R, Raininko R, Tilvis R. White matter changes in healthy elderly persons correlate with attention and speed of mental processing. Arch Neurol. 1993; 50: 818–824.
Scheltens P, Barkhof F, Leys D, Wolters EC, Ravid R, Kamphorst W. Histopathologic correlates of white matter changes on MRI in Alzheimer’s disease and normal aging. Neurology. 1995; 45: 883–888.
Awad IA, Johnson PC, Spetzler RF, Hodak JA. Incidental subcortical lesions identified on magnetic resonance imaging in the elderly. II. Postmortem pathological correlations. Stroke. 1986; 17: 1090–1097.
Braffman BH, Zimmerman RA, Trojanowski JQ, Gonatas NK, Hickey WF, Schlaepfer WW. Brain MR: pathologic correlation with gross and histopathology. 1. Lacunar infarction and Virchow-Robin spaces. AJR Am J Roentgenol. 1988; 151: 551–558.
Braffman BH, Zimmerman RA, Trojanowski JQ, Gonatas NK, Hickey WF, Schlaepfer WW. Brain MR: pathologic correlation with gross and histopathology. 2. Hyperintense white-matter foci in the elderly. AJR Am J Roentgenol. 1988; 151: 559–566.
Marshall VG, Bradley WG Jr, Marshall CE, Bhoopat T, Rhodes RH. Deep white matter infarction: correlation of MR imaging and histopathologic findings. Radiology. 1988; 167: 517–522.
Leifer D, Buonanno FS, Richardson EP Jr. Clinicopathologic correlations of cranial magnetic resonance imaging of periventricular white matter. Neurology. 1990; 40: 911–918.
Chimowitz MI, Estes ML, Furlan AJ, Awad IA. Further observations on the pathology of subcortical lesions identified on magnetic resonance imaging. Arch Neurol. 1992; 49: 747–752.
Munoz DG, Hastak SM, Harper B, Lee D, Hachinski VC. Pathologic correlates of increased signals of the centrum ovale on magnetic resonance imaging. Arch Neurol. 1993; 50: 492–497.
Fazekas F, Kleinert R, Offenbacher H, Schmidt R, Kleinert G, Payer F, Radner H, Lechner H. Pathologic correlates of incidental MRI white matter signal hyperintensities. Neurology. 1993; 43: 1683–1689.
Scarpelli M, Salvolini U, Diamanti L, Montironi R, Chiaromoni L, Maricotti M. MRI and pathological examination of post-mortem brains: the problem of white matter high signal areas. Neuroradiology. 1994; 36: 393–398.
Fazekas F, Kleinert R, Offenbacher H, Payer F, Schmidt R, Kleinert G, Radner H, Lechner H. The morphologic correlate of incidental punctate white matter hyperintensities on MR images. AJNR Am J Neuroradiol. 1991; 12: 915–921.
McKhann G, Drachman D, Folstein M, Katzman R, Price D, Stadlan EM. Clinical diagnosis of Alzheimer’s disease: report of the NINCDS-ADRDA Work Group under the auspices of Department of Health and Human Services Task Force on Alzheimer’s Disease. Neurology. 1984; 34: 939–944.
Graham JE, Rockwood K, Beattie BL, Eastwood R, Gauthier S, Tuokko H, McDowell I. Prevalence and severity of cognitive impairment with and without dementia in an elderly population. Lancet. 1997; 349: 1793–1796.
Jack CR Jr, O’Brien PC, Rettman DW, Shiung MM, Xu Y, Muthupillai R, Manduca A, Avula R, Erickson BJ. FLAIR histogram segmentation for measurement of leukoaraiosis volume. J Magn Reson Imaging. 2001; 14: 668–676.
DeCarli C, Maisog J, Murphy DG, Teichberg D, Rapoport SI, Horwitz B. Method for quantification of brain, ventricular, and subarachnoid CSF volumes from MR images. J Comput Assist Tomogr. 1992; 16: 274–284.
DeCarli C, Murphy DG, Tranh M, Grady CL, Haxby JV, Gillette JA, Salerno JA, Gonzales-Aviles A, Horwitz B, Rapoport SI, et al. The effect of white matter hyperintensity volume on brain structure, cognitive performance, and cerebral metabolism of glucose in 51 healthy adults. Neurology. 1995; 45: 2077–2084.
Bryan RN, Manolio TA, Schertz LD, Jungreis C, Poirier VC, Elster AD, Kronmal RA. A method for using MR to evaluate the effects of cardiovascular disease on the brain: the cardiovascular health study. AJNR Am J Neuroradiol. 1994; 15: 1625–1633.
de Leeuw FE, de Groot JC, Achten E, Oudkerk M, Ramos LM, Heijboer R, Hofman A, Jolles J, van Gijn J, Breteler MM. Prevalence of cerebral white matter lesions in elderly people: a population based magnetic resonance imaging study. The Rotterdam Scan Study. J Neurol Neurosurg Psychiatry. 2001; 70: 9–14.
Wen W, Sachdev P. The topography of white matter hyperintensities on brain MRI in healthy 60- to 64-year-old individuals. NeuroImage. 2004; 22: 144–154.
Schmidt R, Enzinger C, Ropele S, Schmidt H, Fazekas F. Progression of cerebral white matter lesions: 6-year results of the Austrian Stroke Prevention Study. Lancet. 2003; 361: 2046–2048.
Mayer PL, Kier EL. The controversy of the periventricular white matter circulation: a review of the anatomic literature. Am J Neuroradiol. 1991; 12: 223–228.
Roman GC. Senile dementia of the Binswanger type: a vascular form of dementia in the elderly. J Am Med Assoc. 1987; 258: 1782–1788.
Nelson MD, Gonzales-Gomez I, Gilles FH. The search for human telencephalic ventriculofugal arteries. AJNR Am J Neuroradiol. 1991; 12: 215–222.
Moody DM, Bell MA, Challa VR. Features of the cerebral vascular pattern that predict vulnerability to perfusion or oxygenation deficiency: an anatomic study. AJNR Am J Neuroradiol. 1990; 11: 431–439.
Fazekas F, Schmidt R, Scheltens P. Pathophysiologic mechanisms in the development of age-related white matter changes of the brain. Dement Geriatr Cogn Disord. 1998; 9 (Suppl 1): 2–5.
Fazekas F, Schmidt R, Kleinert R, Kapeller P, Roob G, Flooh E. The spectrum of age-associated brain abnormalities: their measurement and histopathological correlates. J Neural Transm Suppl. 1998; 53: 31–39.
DeCarli C, Miller BL, Swan GE, Reed T, Wolf PA, Garner J, Jack L, Carmelli D. Predictors of brain morphology for the men of the NHLBI twin study. Stroke. 1999; 30: 529–536.
Barber R, Gholkar A, Scheltens P, Ballard C, McKeith IG, O’Brien JT. MRI volumetric correlates of white matter lesions in dementia with Lewy bodies and Alzheimer’s disease. Int J Geriatr Psychiatry. 2000; 15: 911–916.
DeCarli C, Miller BL, Swan GE, Reed T, Wolf PA, Carmelli D. Cerebrovascular and brain morphologic correlates of mild cognitive impairment in the National Heart, Lung, and Blood Institute Twin Study. Arch Neurol. 2001; 58: 643–647.
Wu CC, Mungas D, Petkov CI, Eberling JL, Zrelak PA, Buonocore MH, Brunberg JA, Haan MN, Jagust WJ. Brain structure and cognition in a community sample of elderly Latinos. Neurology. 2002; 59: 383–391.
Thomas AJ, Perry R, Kalaria RN, Oakley A, McMeekin W, O’Brien JT. Neuropathological evidence for ischemia in the white matter of the dorsolateral prefrontal cortex in late-life depression. Int J Geriatr Psychiatry. 2003; 18: 7–13.
Coffey CE, Wilkinson WE, Weiner RD, Parashos IA, Djang WT, Webb MC, Figiel GS, Spritzer CE. Quantitative cerebral anatomy in depression. A controlled magnetic resonance imaging study. Arch Gen Psychiatry. 1993; 50: 7–16.
DeCarli C, Murphy DG, Teichberg D, Campbell G, Sobering GS. Local histogram correction of MRI spatially dependent image pixel intensity nonuniformity. J Magn Reson Imaging. 1996; 6: 519–528.
Maes F, Collignon A, Vandermeulen D, Marchal G, Suetens P. Multimodality image registration by maximization of mutual information. IEEE Trans Med Imaging. 1997; 16: 187–198.
Studholme C, Hill DLG, Hawkes DJ. An overlap invariant entropy measure of 3D medical image alignment. Pattern Recognition. 1999; 32: 71–86.
Otte M. Elastic registration of fMRI data using Bezier-spline transformations. IEEE Trans Med Imaging. 2001; 20: 193–206.
Besag J. On the statistical analysis of dirty pictures. Journal Royal Statistical Society B. 1986; 48: 259–302.