aging 148 (2006) 133–142www.elsevier.com/locate/psychresns

Psychiatry Research: Neuroim

A fully automated method for quantifying and localizing whitematter hyperintensities on MR images

Minjie Wua, Caterina Rosanob, Meryl Buttersc, Ellen Whytec, Megan Nablec,Ryan Crooksb, Carolyn C. Meltzerd, Charles F. Reynolds IIIc, Howard J. Aizensteinc,⁎

aDepartment of Electrical and Computer Engineering, University of Pittsburgh, USAbDepartment of Epidemiology, University of Pittsburgh, USAcDepartment of Psychiatry, University of Pittsburgh, USAdDepartment of Radiology, University of Pittsburgh, USA

Received 13 December 2005; received in revised form 16 June 2006; accepted 11 September 2006

Abstract

White matter hyperintensities (WMH), commonly found on T2-weighted FLAIR brain MR images in the elderly, are associatedwith a number of neuropsychiatric disorders, including vascular dementia, Alzheimer's disease, and late-life depression. PreviousMRI studies of WMHs have primarily relied on the subjective and global (i.e., full-brain) ratings of WMH grade. In the currentstudy we implement and validate an automated method for quantifying and localizing WMHs. We adapt a fuzzy-connectedalgorithm to automate the segmentation of WMHs and use a demons-based image registration to automate the anatomiclocalization of the WMHs using the Johns Hopkins University White Matter Atlas. The method is validated using the brain MRimages acquired from eleven elderly subjects with late-onset late-life depression (LLD) and eight elderly controls. This dataset waschosen because LLD subjects are known to have significant WMH burden. The volumes of WMH identified in our automatedmethod are compared with the accepted gold standard (manual ratings). A significant correlation of the automated method and themanual ratings is found (Pb0.0001), thus demonstrating similar WMH quantifications of both methods. As has been shown inother studies (e.g. [Taylor, W.D., MacFall, J.R., Steffens, D.C., Payne, M.E., Provenzale, J.M., Krishnan, K.R., 2003. Localizationof age-associated white matter hyperintensities in late-life depression. Progress in Neuro-Psychopharmacology and BiologicalPsychiatry. 27 (3), 539–544.]), we found there was a significantly greater WMH burden in the LLD subjects versus the controls forboth the manual and automated method. The effect size was greater for the automated method, suggesting that it is a more specificmeasure. Additionally, we describe the anatomic localization of the WMHs in LLD subjects as well as in the control subjects, anddetect the regions of interest (ROIs) specific for the WMH burden of LLD patients. Given the emergence of large NeuroImagedatabases, techniques, such as that described here, will allow for a better understanding of the relationship between WMHs andneuropsychiatric disorders.© 2006 Elsevier Ireland Ltd. All rights reserved.

Keywords: White matter hyperintensity; Late-onset late-life depression

⁎ Corresponding author. Western Psychiatric Institute and Clinic, 3811 O'Hara Street, Pittsburgh, PA 15213, USA. Tel.: +1 412 624 4997; fax: +1412 624 0223.

E-mail address: aizen@pitt.edu (H.J. Aizenstein).

0925-4927/$ - see front matter © 2006 Elsevier Ireland Ltd. All rights reserved.doi:10.1016/j.pscychresns.2006.09.003

134 M. Wu et al. / Psychiatry Research: Neuroimaging 148 (2006) 133–142

1. Introduction

A number of previous studies have shown that whitematter hyperintensities (WMH), also called leuokoar-aiosis, commonly seen on T2-weighted FLAIR MRimages, are associated with neuropsychiatric disorders,including vascular dementia (van Gijn, 1998), Alzhei-mer's disease (Mirsen et al., 1991), and late-onset late-lifedepression (Hickie and Scott, 1998; Thomas et al., 2004).Two analysis strategies have been used to evaluateWMHs on MR brain images: (1) semi-quantitative ratingsystems and (2) quantitative volumetric analyses. Insemi-quantitative system, the WMHs are visually gradedby trained expert raters. The rater assigns eachMR imagea WMH severity score based on its visual similarity to‘prototype’MR images. Typical scales range from low tohigh severity using 4-point or 10-point scales (Fazekaset al., 1987; Bryan et al., 1994; Yue et al., 1997). Thismethod requires subjective judgment; it describes theWMHs through 4 or 10 crude grades. It does not allowaccurate information about the location or volume of theWMHs, and thus may ignore some subtle WMH dif-ferences across groups. Also different visual rating scalesmake it difficult to compare or reproduce the findings onWMHs across centers (Davis et al., 1992).

For the quantitative analyses on WMHs, severalmethods have been explored to automatically or semi-automatically segment theWMHs. For example,K-NearestNeighbor (KNN) classification method was used toautomatically or semi-automatically label the T2-weightedMR brain images as gray matter, CSF and white matterlesions (Kikinis et al., 1992; Swartz et al., 2002; Anbeek etal., 2004b,a). In this method, the classification of an imagevoxel from a new patient relies on the voxel intensities andspatial information of a previously manually classifiedtraining set. Since theMR image of different subjects at thesame center or across centers may have different intensitydistribution ranges, and the normal anatomic variationsacross subjects lead to variability in the spatial features, thismethodmay encounter difficulties for some subjects. Othermachine learning algorithms including artificial neuralnetworks (Pachai et al., 1998) have also been investigatedfor WMH segmentation, which may face similar depen-dencies on a training set. An automated method fromStamatakis is used to delineate large brain lesions on T1-weighted structural images, which involves comparing thesmoothed individual T1-weighted image to a control groupusing general linear model (GLM). The accuracy of thismethod depends on the performance of the spatialnormalization technique. The normal anatomical variationsin brain structure between the individual subject and thecontrol group may present a problem for the registration

accuracy and GLM, so a Gaussian smoothing filter is usedto smooth out the anatomical differences, which may alsoaffect the reliability of the volumetric quantification of thelesions (Stamatakis and Tyler, 2005).

On T2-weighted FLAIR MR images, the WMHsusually have a higher intensity than normal white matter(WM). Some methods automatically or semi-automati-cally segment the WMHs on FLAIR images by defininga cut-off threshold on the images. For example,3.5 standard deviations (S.D.) of the intensity value ofthe normal WM has been used as the lower intensitythreshold for WMH segmentation (Hirono et al., 2000).The histogram of the FLAIR image has been used in aregression model to decide on a cut-off intensity thresh-old, with the pixels above the threshold classified asWMHs (Jack et al., 2001). Another method uses themean and standard deviations of the gray matter, whitematter and CSF to estimate the intensity threshold forWMH, in which a probability map is used to favor themost likely WM regions (Wen and Sachdev, 2004).These methods use only a single intensity threshold tosegment the WMHs for the whole brain or for each sliceof the brain images, which may misclassify some non-WMHs as WMHs, since some gray matter demonstratessignal intensity above the threshold (Hirono et al., 2000),and also the image intensity inhomogeneities may beproblematic. To exclude the misclassified voxels, a man-ually outlined mask of WMHs with surrounding WM,GM and CSF has been used as WMH mask in Hirono'spaper, while inWen's paper a WM probability map (MNI152 brains) has been used to favor the most likely WMregions. Manually outlining the WMH mask of a 3Dbrain volume is time-consuming and labor-intensive,while using a WM probability map in a MNI template tofavor the WM regions in the WMH segmentation of thesubjects will make the segmentation accuracies depen-dent on an accurate inter-subject registration.

Previous research suggests that the location or dis-tribution of WMHs is associated with specific symptoms(Benson et al., 2002). Most previous research focusedonly on WMH visual inspection or volume measurementand did not distinguish anatomically distinct WMHs,while a few groups have explored semi-automated orautomated methods to localize WMHs into large com-partments or categories such as periventricular whitematter hyperintensities (PVWMHs) and deep white mat-ter hyperintensities (DWMHs). For example, in (Swartzet al., 2002), a 3D classification algorithm was applied toseparate DWMHs from PVWMHs. Other investigatorshave used nonlinear image registration methods toconvert the WMHs across subjects into a standardspace (Taylor et al., 2003; DeCarli et al., 2005).

Table 1Clinical characteristics of the subjects

Group I(depressed)

Group II(controls)

t-testprobability

No. of subjects 11 8Age, year (range) 72.2±5.3

(63–80)72.3±4.8(67–81)

0.93560809

Gender, M/F 5/6 4/4WMH scores±S.D. 2.55±1.9 1.25±0.5 0.05412192MMSE±S.D. 27.7±3.6 28.8±1.5 0.40560931Hamilton±S.D. 20.3±4.9 2±2.07 2.2477E−08MATTIS±S.D. 136.3±5.9 139.9±3.4 0.11628563

Statistical comparisons utilized a two-sample, unequal variance, two-tailed Student's t-test.

135M. Wu et al. / Psychiatry Research: Neuroimaging 148 (2006) 133–142

In the current study, we present an alternative auto-mated method for WMH quantification and localization,which uses a fuzzy-connected algorithm to segment theWMHs, and the Automated Labeling Pathway (ALP) tolocalize the WMHs into the anatomical space (Wu et al.,2006). Previous research used fuzzy-connected algo-rithm for semi-automated WMH segmentation (Miki etal., 1997; Udupa et al., 1997), which required some userinteraction and did not give spatial information on theWMHs. Our automated method uses the histogram ofthe FLAIR image to automatically generate the WMHseeds, and then the fuzzy-connected algorithm usesspecific parameters to form a WMH cluster (containingthe respective seed). The system updates the seedsiteratively and combines the scattered WMH clustersinto the final WMH segmentation. Since the fuzzy-connected algorithm uses different parameters for eachseed, this method enables different threshold for eachWMH cluster and avoids a single cut-off threshold forthe whole brain or brain slice. This potentially offersmore precise WMH segmentation. The method auto-matically identifies WMH seeds and generates WMHsegmentation, which is objective and does not requireany manual interaction. A fully deformable registration(ALP; Wu et al., 2006), which combines the piecewiselinear registration for coarse alignment with Demonsalgorithm for voxel-level refinement, is used for accu-rate WMH localization on the Johns Hopkins UniversityWhite Matter Atlas (Wakana et al., 2004).

We report the results of a quantitative assessmentWMH of a group of elderly control subjects compared toa group of LLD subjects. This group was chosen becauseit is known that these subjects have a high WMH burden(O'Brien et al., 1996). We compare the WMH volumesidentified with our approach to the gold standardassessments based on manual expert ratings. Addition-ally, the anatomical localization of the WMHs found withour approach is described, and the WMH burden of thecontrol group is region-wise statistically compared to thatof the LLD patient group.

2. Materials and methods

2.1. Subjects

The 19 subjects (eleven patients and eight controls)were recruited through the University of PittsburghIntervention Research Center for Late-Life Mood Dis-orders. Subjects were 63 to 81 years of age (mean age=72.3, S.D.=4.86), whose WMH visual scores rangedfrom 0.5 to 6.5 (mean WMH score=2, S.D.=1.6). Allsubjects (controls and depressed) received a SCID-IV

evaluation, which was reviewed in a diagnostic consen-sus conference. Eleven of the 19 subjects were diagnosedas depressed patients; while the remaining eight subjectswere termed control subjects.

The 11 patients had late-onset late-life depression;they met DSM-IV criteria for Major Depressive Disorder(American Psychiatric Association, 2000) and theirdepression began at the age of 60 years or older. Themean Hamilton Depression Rating Scale on patients was20.3 (S.D.=4.9). The subjects did not have significantcognitive impairment; mean Mattis Dementia RatingScale was 136.3 (S.D.=5.9). They were all participants ina research trial of antidepressant medications. Other thanMajor Depressive Disorder (for subjects in the depressedgroup) and anxiety disorders, all other Axis I psychiatricdisorders were used as exclusion criteria. We chose toinclude subjects with co-morbid anxiety disorders due tothe high prevalence (48%) of anxiety disorders insubjects with late-life depression (Beekman et al., 2000).

Each subject was assessed by the Mini Mental StateExamination (MMSE), Hamilton Rating Scale for De-pression (Hamilton), and Mattis Dementia Rating Scale(Mattis). Clinical characteristics of the subjects (patientsand controls) are summarized in Table 1. The 2 groupswere well balanced with respect to gender and age.

The MR images used in the current analyses wereobtained at the time of subject enrollment, before theantidepressant medication was started. This study wasapproved by the University of Pittsburgh InstitutionalReview Board (IRB). Written informed consent wasobtained.

2.2. MR imaging parameters

Magnetic resonance images were acquired on a 1.5 TSigna Scanner (GE Medical Systems, Milwaukee, WI).The 3D structural MR images were acquired at sagittalorientation using 3D Spoiled GRASS (SPGR, TR/TE=

Fig. 1. WMH segmentation flowchart. The processing steps used to automatically segment the WMHs on FLAIR MR brain images.

136 M. Wu et al. / Psychiatry Research: Neuroimaging 148 (2006) 133–142

5/25 ms; flip angle=40°; FOV=24×18 cm2, slicethickness=1.5 mm, matrix=256×192 matrix).

The following axial series were also obtained: T1-weighted (TR/TE=500/11 ms, NEX=1); fast fluid-attenuated inversion recovery (fast FLAIR) (TR/TE=9002/56 ms Ef; TI=2200 ms, NEX=1). Section thick-ness was 5 mm with a 1-mm inter-section gap. All axialsequences were obtained with a 24 cm field of view and a192×256 pixel matrix. Slice thickness and orientation

Fig. 2. ALP flowchart. The processing steps that constitute our automated labestimates. The process uses a variety of publicly available packages, as well asimages.

were chosen so that the acquired images would becompatible with theWMH rating scales described below.

2.3. White matter hyperintensity ratings

The WMH ratings were based on a system developedfor the Cardiovascular Health Study (CHS; Bryan et al.,1994; Yue et al., 1997). A numerical rating for theWMHswas assigned by the comparison of each subject's

eling pathway (ALP), which is used to generate regional brain volumesome locally developed programs, for atlas-based segmentation of MR

137M. Wu et al. / Psychiatry Research: Neuroimaging 148 (2006) 133–142

imaging data to predefined CHS visual standards andrepresentative of progressive severity within a 10-pointscale (0 through 9). Two raters independently evaluatedWMH on the FLAIR images. If they differed in theirratings by one point, the final rating was the mean of thetwo values. A greater than 1-point difference betweenraters was considered as a disagreement, and was adju-dicated by consensus.

2.4. Automated WMH segmentation and localization

The major steps of the automated WMH segmentationprocedure involved (1) image preprocessing, (2) auto-mated WMH segmentation, and (3) automated WMHlocalization. Image preprocessing included skull strip-ping of the SPGR and FLAIR brain images, which im-proved the accuracies of WMH segmentation andlocalization. For the skull stripping on the FLAIR im-ages, we used the Brain Extraction Tool (BET, Smith,2002) on the T1-weighted images, which were acquiredat the same location and voxel-size as the FLAIR images.The resulting stripped T1-weighted image was then usedas a brain mask to remove the skull and scalp from theFLAIR image.

This automated WMH segmentation method in-volved four steps: (1) automatically identifying WMHseeds based on the intensity histogram of the FLAIRimage, (2) using a fuzzy-connected algorithm to seg-ment the WMH clusters, (3) iteratively updating the set

Fig. 3. An overview of the WM

of seeds, and (4) combining the WMH clusters into thefinal WMH segmentation. The histogram of the skull-stripped FLAIR image was used to define a threshold(mean±3 S.D.) for seed selection; voxels beyond thisthreshold were classified as WMHs, which were used asseeds in the fuzzy-connected algorithm to segment sur-rounding WMH voxels. The background of the FLAIRimage was excluded when calculating its intensityhistogram, mean intensity and standard deviation. Inthe fuzzy-connected algorithm, the fuzzy adjacency andaffinity, both between 0 and 1, are defined for each pairof voxels (a,b): the fuzzy adjacency μα(a,b) defines howclose the two voxels are, while the affinity μk(a,b)(determined based on adjacency degree μα(a,b) andintensity similarity) indicates how strongly the twovoxels “hang together” in space and intensity. A fuzzy-connected object is a set of voxels O with properties asfollows: any two voxels (a,b) from O have an affinity μk(a,b)Nx, 0≤x≤1, and for any pair a∈O, b∉O, theaffinity μk(a,b)bx,0≤ x≤1, a detailed and precisemathematics definition is given in Udupa and Samar-asekera (1996) and Udupa et al. (1997). For eachselected WMH seed the fuzzy-connected algorithmgenerates a fuzzy object, within which each pair ofvoxels has a strong fuzzy-connectedness or affinity(above certain threshold, 0.5 in this study), and thesystem automatically delineates a 3D WMH clustercontaining the respective seed. Multiple 3D FLAIRimage WMH clusters are generated from the set of

H localization procedure.

Table 2Mean volumes of WMH (mm3) per region for the control group(8 controls) versus the patient group (11 patients) and the t-test resultson normalized WMHs

Region/Group

Mean WMH volume (mm3) Two-tailedt-test onnormalizedWMH

Control Patient

Whole brain 2737.3437 8541.030942 0.042665436ATRL 649.7712 2211.418354 0.055444463ATRR 821.34 1764.802473 0.018656432CCF 201.6846 1165.593838 0.016828416CCO 416.1456 1529.608106 0.012737393CSTL 148.2975 309.167762 0.057927493CSTR 158.3361 652.5467107 0.060871669CgLL 35.1351 110.2963835 0.081089283CgLR 8.6697 40.42591736 0.060644356CgUL 11.8638 550.094162 0.136364206CgUR 24.1839 29.50488595 0.818743765IFOL 338.1183 1226.836086 0.034295001IFOR 478.6587 2039.035002 0.039086461ILFL 192.5586 716.5644694 0.018575153ILFR 214.9173 860.4083306 0.038395966SLFBL 273.3237 929.735762 0.35930778SLFBR 156.9672 1441.153785 0.300544506SLFTL 128.6766 717.4695273 0.405860665SLFTR 46.5426 761.4553388 0.297697885UNCL 104.0364 430.2645025 0.168513136UNCR 139.1715 327.8119537 0.030380464

Keys: ATRL/R – anterior thalamic radiation (left or right), CCF/O –corpus callosum (frontal or occipital), CSTL/R – corticospinal tract(left or right), CgLL/R – cingulate (lower part left or right), CgUL/R –cingulate (upper part left or right), IFOL/R – interior fronto-occipitalfasciculus (left or right), ILFL/R – inferior longitudinal fasciculus (leftor right), SLFBL/R – entire superior longitudinal fasciculus (left orright), SLFTL/R – superior longitudinal fasciculus (the branch to thetemporal lobe, left or right), UNCL/R – uncinate fasciculus (left orright).

138 M. Wu et al. / Psychiatry Research: Neuroimaging 148 (2006) 133–142

automatically selected seeds and then combined to forman overall WMH segmentation volume.

The flow chart of the WMH segmentation is shown inFig. 1. The fully automated WMH segmentation systemwas implemented in C++ and ITK. The WMH seg-mentation algorithm is available to the readers uponrequest through our website (http://www.pitt.edu/~aizen/GPN_Home.html).

Automated Labeling Pathway (ALP, see Fig. 2) is anautomated method we developed in a series of func-tional and structural MRI studies to automatically labelspecific anatomic regions of interest (Rosano et al.,2005; Aizenstein et al., 2005; Wu et al., 2006). Thepathway combines a series of publicly available soft-ware packages such as AFNI (Cox, 1996), BET (Smith,2002), FLIRT (Jenkinson et al., 2002) and ITK (Yoo,2004), as well as some locally developed programs toimplement atlas-based segmentation of MR images.ALP is used to automatically label ROIs on the SPGRimage of a subject.

In ALP described in Wu et al. (2006), the inter-subjectregistration (template colin27→subject 3D SPGR) isdone using a fully deformable registration model similarto that described by Chen (1999). We have implementedthis using the registration library in Insight Segmentationand Registration Toolkit (ITK, Yoo, 2004). This methodstarts with a grid-based piecewise linear registration andthen uses a demons registration algorithm as a fine-tuningprocedure for a voxel-level spatial deformation. The fullydeformable registration allows for a high degree of spatialdeformation, which seems to give it a particularadvantage over other standard registration packages,such as Automated Image Registration (AIR) andStatistical Parametric Mapping (SPM).

An overview of WMH localization procedure issummarized in Fig. 3. The high-resolution referenceimage (MNI colin27) is registered to the T1-weightedSPGR high-resolution image of the subject using ALP,and the Johns Hopkins University White Matter Atlas(defined on the reference brain MNI colin27 image) iswarped into each individual's anatomic image space.Then the anatomic information in subject SPGR space istransformed further into the subject's FLAIR image spaceby rigid-body registration between the subject SPGRimage and subject T1 in-plane image, which was ac-quired the same slice prescription as the subject's FLAIRimage. In this way, the anatomical information in the atlasis carried into the subject's FLAIR space and the ROIslabeled on the subject's FLAIR image are used as binarymasks to localize the WMHs. In this procedure, theWMH localization task is viewed as a registrationprocedure. The Johns Hopkins University White Matter

Atlas we used in the current study is based on high-resolution diffusion tensor MR imaging and 3d tractreconstruction. The atlas has 17 prominent white tractsincluding anterior thalamic radiation (ATR), cingulum(Cg) and other tracts (Wakana et al., 2004), as listed inTable 2.

Prior to the ALP registration procedure, the non-brain tissues such as skull and scalp are stripped fromthe subject's 3D SPGR image using BET (Smith, 2002).A simple morphological method involving threshold,erosion, dilation and hole-filling is used to improve theskull stripping result (Wu et al., 2005). Also a rigidalignment of the anterior and posterior commissures(AC–PC) and intensity normalization are done on eachsubject's 3D SPGR image as well as on the templatecolin27, which gives each subject the same orientationand image intensity distribution as the template, andtherefore improves the registration accuracy.

139M. Wu et al. / Psychiatry Research: Neuroimaging 148 (2006) 133–142

3. Results and discussion

A subject with some discrete lesions (as well asconfluent lesions) is chosen to demonstrate the results ofthis WMH extraction algorithm. Nine pairs of the seg-mentedWMH slices versus corresponding FLAIR slicesfrom the subject are displayed in Fig. 4, showing thismethod's effectiveness in the segmentation of discreteas well as confluent WMHs.

3.1. WMH segmentation evaluation

The WMH segmentation results of 19 subjects usingthis automated method were statistically compared to theWMH visual grades from the manual ratings. The com-parison was done with a linear regression model. In thisstudy we chose to use semi-quantitative CHS ratings asthe gold standard for comparison. An alternative approachwould have been to use manually segmented WMHtracings. Since the two measures being compared useddifferent metrics, we are only demonstrating a correlationbetween the measures rather an absolute agreement.

The WMH volumes of the 19 subjects from the auto-mated segmentation method were found to be signifi-cantly correlated to the visual grades with a R2=0.909and F(1,18)=170.7, Pb0.0001. Since the visual grade is

Fig. 4. Automated WMH segmentation results on the FLAIR MR images ofeach paired slices, the left slice is the FLAIR slice and the right one is the a

a global index to the WMH severity on the subject brainimage, the WMH volume is normalized by the overallWM volume (calculated from SPGR brain image). Thenormalized WMH results were also significantly corre-lated to the visual grades [R2=0.909, F(1,18)=170.3,Pb0.0001]. This WM normalization method may not bethe best way for whole brain adjustment, since previousstudies have showed that WMH are significantly relatedto atrophy (Capizzano et al., 2004; Schmidt et al., 2005).A whole brain normalization method, which takes brainatrophy into consideration, may be better for WMHassessment.

The high correlation between the normalized WMHquantifications from the automated method and thevisual grades demonstrates that this automated methodcan successfully segment the WMHs on MR FLAIRimages.

3.2. Localization of WMHs

Using ALP, the Johns Hopkins University WhiteMatter Atlas was transferred to subject's 3D SPGRimage and further carried into subject FLAIR imagespace. The atlas regions in the subjects' FLAIR imagespace were then used as ROI masks to localize theWMHs. Fig. 5 shows the segmented ROIs in MNI

one subject. Nine paired image slices on the subject are shown here. Inssociated automated WMH segmentation result.

140 M. Wu et al. / Psychiatry Research: Neuroimaging 148 (2006) 133–142

template colin27 space, the individual SPGR structuralspace and FLAIR image spaces; respective MR imagesare also shown as underlay images.

The localized WMH volumes were quantified bymultiplying voxel size by the number of WMH voxelsinside the ROIs including anterior thalamic radiations,corticospinal tracts, etc., as listed in Table 2. The WMHvolume estimates from WMH localization describe thespatial distribution of the WMH burden, which can fa-cilitate further research on the role of WMH in patho-genesis of neuropsychiatric disorders. In Table 2, for eachregion of interest, the WMH volumes of the LLD patientgroup were statistically compared to the WMH volumesof the control group using two-tailed two-sample unequalvariance t-test. We found a significant difference inwhole brain WMH volume between the LLD patientgroup and the control group; however, the results fromthe WMH localization method provide more anatomicalspecificity. As shown in Table 2, there was significantdifference in WMH spatial distribution between LLDpatient group and control group in regions including rightanterior thalamic radiation, corpus callosum (CC), in-

Fig. 5. The result of atlas-based segmentation from ALP. Segmentation resultsin the bottom row. (a) The MNI template colin 27, overlapped with the John Hcorpus callosum, corticospinal tract, inferior fronto-occipital, inferior lonfasciculus, etc. (b) A single subject 3d SPGR image, overlapped with the tranwith the transformed ROIs.

ferior fronto-occipital (IFO), inferior longitudinal fascic-ulus (ILF), and right uncinate fasciculus (UNC), while nosignificant difference was found in cingulum (CgLL,CgLR, CgUL, CgUR) and superior longitudinal fascic-ulus (SLFBL, SLFBR, SLFTL, SLFTR).

The current study is limited by the low-resolutionFLAIR image, as well as the limited number of subjects(11 patients and 8 control subjects). The analyzed FLAIRimages were acquired with a slice thickness of 5 mm anda 1 mm gap, which may be an inadequate resolution foraccurate volumetric quantification of the WMHs, whichaccordingly may affect the reliability of the group com-parison results. A higher image resolution, such as a slicethickness of 2 mm and with no gap, could improve theWMH quantification, and improve the registration accu-racy, which would lead to more accurate WMH locali-zation. Also a larger group of well-characterized LLDsubjects with a matched elderly non-depressed controlgroup would add confidence to the WMH localizationfindings in Table 2 that are specific for LLD.

The WMH segmentation and localization method wedescribed provides more specific and more accurate

are shown at axial orientation in the top row and the coronal orientationopkins University White Matter Atlas (i.e., anterior thalamic radiation,gitudinal fasciculus, superior longitudinal fasciculus, right uncinatesformed ROIs. (c) The same single subject FLAIR image, overlapped

141M. Wu et al. / Psychiatry Research: Neuroimaging 148 (2006) 133–142

information about WMH volume and spatial distributionthan visual WMH grades. Also the fully automatedmethod is objective and it does not require any manualinterventions. Unlike different visual grading systems, itis very easy to compare the WMH findings from thismethod across different centers. Themethod relies on theproperties of subject's own FLAIR image such as theintensity distribution of WMHs, the connectivity and thediffusivity of the WMHs for the WMH segmentation,which does not rely on any training dataset as do some ofthe reviewed methods (Kikinis et al., 1992; Swartz et al.,2002; Anbeek et al., 2004b,a).

4. Conclusion

In this report we presented and validated a newmethod for fully automated segmentation and localiza-tion of WMHs on MR images. The method adapts thefuzzy-connected algorithm for WMH segmentation anduses a demons-based fully deformable registration forWMH localization. The automated WMH segmentationmethod was evaluated by comparing the resultingWMHquantifications (non-normalized or normalized by totalWM volumes) of the 19 elderly subjects (11 late-lifedepressed subjects and 8 elderly controls) with the stan-dard visual grading approach for estimating WMHburden. In the comparisons we found a high correlationof the WMH ratings between our new semi-automatedapproach and the manual ratings. Specifically, the twomethods correlate with R2 =0.909, Pb0.0001. Furtherlocalization of WMH follows the expected patterns ofLLD, i.e., high WMH burden in the subcortical, andfrontal regions.

Quantification and localization of WMH volumes iscritical for research into the risk factors and pathogen-esis of neuropsychiatric disorders. Most previousmethods were labor-intensive, subjective, and providedlittle if any anatomic localization. The current methodsolves many of the previous limitations: it does notrequire any manual intervention, provides WMHvolume estimates, and localizes the WMH burden to anumber of anatomic ROIs. Methods such as describedhere are particularly relevant given the emergence oflarge MRI databases, such as that provided by theAlzheimer's Disease Neuroimaging Initiative (http://www.loni.ucla.edu/ADNI/).

The development and implementation of an automat-ed method for quantifying and localizing WMH willfacilitate further, fine-grained understanding of: 1) short-and long-term treatment response; 2) evolution of cog-nitive functioning in late life depression; 3) evolution ofleukoarisosis in LLD; 4) impact of medical and psy-

chiatric treatment on WMHs in LLD; and 5) modeling ofcognitive impairment in LLD: e.g., is diminution in speedof information processing driven primarily by WMH,beta amyloid deposition, or both?

Acknowledgements

This work was supported by NIH grants MH64678,MH37869, MH043832, MH067710, P30MH52247,P30MH71944, P30AG024827, and NARSAD.

References

Aizenstein, H.J., Butters, M.A., Figurski, J.L., Stenger, V.A., Reynolds,C.F.I., Carter, C.S., 2005. Prefrontal and striatal activation duringsequence learning in geriatric depression. Biological Psychiatry 58(4), 290–296.

American Psychiatric Association, 2000. Diagnostic and StatisticalManual of Mental Disorders DSM-IV-TR. American PsychiatricAssociation.

Anbeek, P., Vincken, K.L., van Osch, M.J., Bisschops, R.H., van derGrond, J., 2004a. Automatic segmentation of different-sized whitematter lesions by voxel probability estimation. Medical ImageAnalysis 8 (3), 205–215.

Anbeek, P., Vincken, K.L., van Osch, M.J., Bisschops, R.H., van derGrond, J., 2004b. Probabilistic segmentation of white matterlesions in MR imaging. NeuroImage 21 (3), 1037–1044.

Beekman, A.T., de Beurs, E., van Balkom, A.J., Deeg, D.J., van Dyck,R., van Tilburg, W., 2000. Anxiety and depression in later life: co-occurrence and communality of risk factors. American Journal ofPsychiatry 157 (1), 89–95.

Benson, R.R., Guttmann, C.R.G., Wei, X., Warfield, S.K., Hall, C.,Schmidt, J.A., Kikinis, R.I.W.L., 2002. Older people with impairedmobility have specific loci of periventricular abnormality on MRI.Neurology 58, 48–55.

Bryan, R., Manolio, T., Schertz, L., Jungreis, C., Poirier, V., Elster,A., Kronmal, H., 1994. A method for using MR to evaluate theeffects of cardiovascular disease on the brain: the Cardiovas-cular Health Study. American Journal of Neuroradiology 15,1625–1633.

Capizzano, A.A., Ación, L., Bekinschtein, T., Furman, M., Gomila, H.,Martínez, A., Mizrahi, R., Starkstein, S.E., 2004. White matterhyperintensities are significantly associated with cortical atrophyin Alzheimer's disease. Journal of Neurology, Neurosurgery andPsychiatry 75, 822–827.

Chen, M., 1999. 3-D Deformable Registration Using a Statistical Atlaswith Applications in Medicine. Carnegie Mellon University.

Cox, R.W., 1996. AFNI:software for analysis and visualization offunctional magnetic resonance neuroimages. Computers andBiomedical Research 29, 162–173.

Davis, P.C., Gray, L., Albert, M.S., Wilkinson, W., Hughes, J.,Heyman, A., Gado, M., Kumar, A.J., Destian, S., Lee, C., 1992.The consortium to establish a registry for Alzheimer's disease(CERAD): Part III. Reliability of a standardized MRI evaluation ofAlzheimer's disease. Neurology 42 (9), 1676–1680.

DeCarli, C., Fletcher, E., Ramey, V., Harvey, D., Jagust, W.J.,2005. Anatomical mapping of white matter hyperintensities(WMH): exploring the relationships between periventricularWMH, Deep WMH, and total WMH burden. Stroke 36,50–55.

142 M. Wu et al. / Psychiatry Research: Neuroimaging 148 (2006) 133–142

Fazekas, F., Chawluk, J., Alavi, A., Hurtig,H., Zimmerman, R., 1987.MRsignal abnormalities at 1.5 T in Alzheimer's dementia and normalaging. American Journal of Roentgenology 149 (2), 351–356.

Hickie, I., Scott, E., 1998. Late-onset depressive disorders: apreventable variant of cerebrovascular disease? PsychologicalMedicine 28 (5), 1007–1013.

Hirono, N., Hajime, K., Kazui, H., Hashimoto, M., Mori, E., 2000.Impact of white matter changes on clinical manifestation ofAlzheimer's disease a quantitative study. Stroke 31, 2182–2188.

Jack Jr., C.R., O'Brien, P.C., Rettman, D.W., Shiung, M.M., Xu, Y.,Muthupillai, R., Manduca, A., Avula, R., Erickson, B.J., 2001.FLAIR histogram segmentation for measurement of leukoaraiosisvolume. Journal of Magnetic Resonance Imaging 14 (6),668–676.

Jenkinson, M., Bannister, P., Brady, M., Smith, S., 2002. Improvedoptimization for the robust and accurate linear registration andmotion correction of brain images. Neuroimage 17, 825–841.

Kikinis, R., Shenton, M.E., Gerig, G., Martin, J., Anderson, M.,Metcalf, M., Guttman, C.R., McCarley, R.W., Lorenson, W., Cline,H., et al., 1992. Routine quantitative analysis of brain andcerebrospinal fluid spaces with MR imaging. Journal of MagneticResonance Imaging 2, 619–629.

Miki, Y., Grossman, R.I., Udupa, J.K., Samarasekera, S., van Buchem,M.A., Cooney, B.S., Pollack, S.N., Kolson,D.L., Constantinescu, C.,Polansky, M., et al., 1997. Computer-assisted quantitation ofenhancing lesions in multiple sclerosis: correlation with clinicalclassification. American Journal of Neuroradiology 18 (4), 705–710.

Mirsen, T., Lee, D., Wong, C., Diaz, J., Fox, A., Hachinski, V.,Merskey, H., 1991. Clinical correlates of white-matter changes onmagnetic resonance imaging scans of the brain. Archives ofNeurology 48, 1015–1021.

O'Brien, J., Ames, D., Schwietzer, I., 1996. White matter changes indepression and Alzheimer's disease: a review of magnetic reso-nance imaging studies. International Journal of Geriatric Psychi-atry 11, 681–694.

Pachai, C., Zhu, Y.M., Grimaud, J., Hermier, M., Dromigny-Badin, A.,Boudraa, A., Gimenez, G., Confavreux, C., Froment, J.C., 1998. Apyramidal approach for automatic segmentation of multiple sclerosislesions in brain MRI. Computerized Medical Imaging and Graphics22, 399–408.

Rosano, C., Becker, J., Lopez, O., Lopez-Garcia, P., Carter, C.,Newman, A., Kullere, L., Aizenstein, H., 2005. Morphometricanalysis of gray matter volume in demented older adults: exploratoryanalysis of the Cardiovascular Health Study brain MRI database.Neuroepidemiology 24 (4), 221–229.

Schmidt, R., Ropele, S., Enzinger, C., Petrovic, K., Smith, S., Schmidt,H., Matthews, P.M., Fazekas, F., 2005. White matter lesion pro-gression, brain atrophy, and cognitive decline: the Austrian strokeprevention study. Annals of Neurology 58 (4), 610–616.

Smith, S.M., 2002. Fast robust automated brain extraction. HumanBrain Mapping 17 (3), 143–155.

Stamatakis, E.A., Tyler, L.K., 2005. Identifying lesions on structuralbrain images – validation of the method and application toneuropsychological patients. Brain and Language 94, 167–177.

Swartz, R.H., Black, S.E., Feinstein, A., Rockel, C., Sela, G., Gao, F.Q.,Caldwell, C.B., Bronskill, M.J., 2002. Utility of simultaneousbrain, CSF and hyperintensity quantification in dementia. Psychi-atry Research. Neuroimaging 116, 83–93.

Taylor, W.D., MacFall, J.R., Steffens, D.C., Payne, M.E., Provenzale,J.M., Krishnan, K.R., 2003. Localization of age-associated whitematter hyperintensities in late-life depression. Progress in Neuro-Psychopharmacology & Biological Psychiatry 27 (3), 539–544.

Thomas, A.J., Kalaria, R.N., O'Brien, J.T., 2004. Depression andvascular disease: what is the relationship? Journal of AffectiveDisorders 79 (1–3), 81–95.

Udupa, J.K., Samarasekera, S., 1996. Fuzzy connectedness and objectdefinition: theory, algorithms, and applications in image segmen-tation. Graphical Models and Image Processing 58 (3), 246–261.

Udupa, J.K., Wei, L., Samarasekera, S., Miki, Y., Van Buchem, M.,Grossman, R., 1997. Multiple sclerosis lesion quantification usingfuzzy-connectedness principles. IEEE Transactions on MedicalImaging 16 (5), 598–609.

van Gijn, J., 1998. Leukoaraiosis and vascular dementia. Neurology 51(3 Suppl 3), S3–S8.

Wakana, S., Jiang, H., Nagae-Poetscher, L.M., van Zijl, P.C.M., Mori,S., 2004. Fiber tract-based atlas of human white matter anatomy.Radiology 230, 77–87.

Wen, W., Sachdev, P., 2004. The topography of white matterhyperintensities on brain MRI in healthy 60- to 64-year-oldindividuals. NeuroImage 22 (1), 144–154.

Wu, M., Rosano, C., Aizenstein, H.J., 2005. A morphological methodto improve skull stripping of MR images. Presented at the 11thAnnual Meeting of the Society for Human Brain Mapping,Toronto (Available on CD-ROM from Neuroimaging).

Wu, M., Carmichael, O., Carter, C.S., Figurski, J.L., Lopez-Garcia, P.,Aizenstein, H.J., 2006. Quantitative comparison of neuroimageregistration by AIR, SPM, and a fully deformable model. HumanBrain Mapping 27 (9), 747–754.

Yoo, T., 2004. Insight into Images: Principles and Practice forSegmentation, Registration, and Image Analysis. AK Peters Ltd.,Wellesey, MA.

Yue, N., Arnold, A., Longstreth, W., Elster, A., Jungreis, C., O'Leary,D., Poirier, V., Bryan, R., 1997. Sulcal, ventricular, and whitematter changes at MR imaging in the aging brain: data from theCardiovascular Health Study. Radiology 202, 33–39.