NeuroImage 55 (2011) 566–573

Contents lists available at ScienceDirect

NeuroImage

j ourna l homepage: www.e lsev ie r.com/ locate /yn img

The effect of template selection on diffusion tensor voxel-based analysis results

Wim Van Hecke a,b,c,⁎, Alexander Leemans d, Caroline A. Sage b, Louise Emsell e,f, Jelle Veraart c, Jan Sijbers c,Stefan Sunaert b, Paul M. Parizel a

a Department of Radiology, University Hospital Antwerp, Edegem (Antwerp), Belgiumb Department of Radiology, University Hospital Leuven, Leuven, Belgiumc Department of Physics, Visionlab, University of Antwerp, Wilrijk (Antwerp), Belgiumd Image Sciences Institute, University Medical Center Utrecht, Utrecht, the Netherlandse Department of Developmental and Functional Brain Imaging, Murdoch Children's Research Institute, Melbourne, Australiaf Department of Psychiatry, National University of Ireland Galway, Galway, Ireland

⁎ Corresponding author. Department of Radiology,Wilrijkstraat 10, B-2650 Antwerp, Belgium.

E-mail address: wim.vanhecke@ua.ac.be (W. Van He

1053-8119/$ – see front matter © 2010 Elsevier Inc. Aldoi:10.1016/j.neuroimage.2010.12.005

a b s t r a c t

a r t i c l e i n f o

Article history:Received 17 June 2010Revised 3 November 2010Accepted 2 December 2010Available online 10 December 2010

Keywords:Diffusion tensor imagingVoxel-based analysisAtlasTemplate

Diffusion tensor imaging (DTI) is increasingly being used to study white matter (WM) degeneration inpatients with psychiatric and neurological disorders. In order to compare diffusion measures across subjectsin an automated way, voxel-based analysis (VBA) methods were introduced. In VBA, all DTI data aretransformed to a template, after which the diffusion measures of control subjects and patients are comparedquantitatively in each voxel. Although VBA has many advantages compared to other post-processingapproaches, such as region of interest analysis or tractography, VBA results need to be interpreted cautiously,since it has been demonstrated that they depend on the different parameter settings that are applied in theVBA processing pipeline.In this paper, we examine the effect of the template selection on the VBA results of DTI data. We hypothesizedthat the choice of template to which all data are transformed would also affect the VBA results. To this end,simulated DTI data sets as well as DTI data from control subjects and multiple sclerosis patients were alignedto (i) a population-specific DTI template, (ii) a subject-based DTI atlas in MNI space, and (iii) the ICBM-81 DTIatlas. Our results suggest that the highest sensitivity and specificity to detect WM abnormalities in a VBAsetting was achieved using the population-specific DTI atlas, presumably due to the better spatial imagealignment to this template.

Antwerp University Hospital,

cke).

l rights reserved.

© 2010 Elsevier Inc. All rights reserved.

Introduction

Recently, voxel-based analysis (VBA) studies have demonstratedthe potential of diffusion tensormagnetic resonance imaging (DT-MRIor DTI) to detect white matter (WM) changes in patients with variousneurological or psychiatric disorders (White et al., 2007; Sundgrenet al., 2004). In VBA, all DTI data sets are first transformed to an atlas ortemplate (Mori et al., 2009). Subsequently, the diffusion measures ofcontrol subjects and patients are compared in each voxel (Ashburnerand Friston, 2000). Although this VBA approach has many advantagescompared to other post-processing methods, such as the region ofinterest (ROI) analysis, VBA results should be interpreted cautiously,since VBA results have been shown to depend on the selection ofdifferent settings in the VBA processing pipeline, such as thecoregistration method, smoothing kernel width, statistics, and post-hoc analysis (Jones et al., 2005; Smith et al., 2006; Zhang et al., 2007;Hsu et al., 2008, 2010; Sage et al., 2009; Van Hecke et al., 2010a). For

example, since in VBA, the statistical tests are performed on a voxellevel, it is important that spatially overlapping voxels of differentsubjects correspond to the same anatomical structure after imagealignment to the atlas. In this context, it has already been reportedthat the VBA results depend on the image coregistration algorithmthat is used in the analysis (Zhang et al., 2007; Sage et al., 2009).

In most VBA studies of DTI data, a standard template, such as theMontreal Neurological Institute (MNI) atlas, is used as a referencespace for the alignment of all DTI data sets. The advantage of thistemplate is that it contains anatomic and cytoarchitectonic labels in astandard coordinate space, allowing standardized reporting andcomparison across studies. However, since this atlas is not study-specific, it might fail to provide a good representation of the groupthat is examined (e.g. brain structures of neonatal or older subjectscan differ from the structures in the MNI atlas), thereby potentiallyresulting in considerable residual image misalignment after coregis-tration of data sets to MNI space. Furthermore, as the original MNItemplate only contains anatomical MR information, many studies usethe T1 or T2 image intensities as input information for thecoregistration algorithm. In other studies, the deformation field thatwas obtained from the coregistration of the anatomical MR image to

Fig. 1. The 19 locations of simulated white matter degeneration are shown in differentcolors and indicated by a white arrow on axial slices of the FA map.

567W. Van Hecke et al. / NeuroImage 55 (2011) 566–573

MNI space is subsequently applied to the fractional anisotropy (FA)map of the same subject to transform DTI data to MNI space(Kyriakopoulos et al., 2007). This can affect the reliability of the VBAresults, since it has been demonstrated that the use of the multi-valued tensor information significantly improves the alignment of DTIdata sets (Park et al., 2003; Van Hecke et al., 2007). Recently, Moriet al. (2008) introduced a stereotaxic WM template (the ICBM-81atlas) that was constructed from 81 DTI data sets of healthy subjectsthat were normalized with an affine transformation to the ICBM-152template (Mori et al., 2008). This template contains the tensorelements, so that they can be used in a multi-channel coregistrationapproach. In other VBA studies, a single subject data set of the imagegroup was selected as the template image (Jones et al., 2002; Smithet al., 2006; Douaud et al., 2007). Although such an atlas can beregarded as study specific, it might fail to be a good representative ofthe whole subject group, especially when pathology is involved.

In order to copewith the aforementioned issues in the construction ofa DTI template for VBA, we recently proposed a population-based, study-specific DTI atlas, in which the magnitudes of the deformation fields thatare needed to transform the different DTI data sets of the study group tothe atlas are minimized, leading to a decreased image misalignment andmore reliableVBA results (VanHecke et al., 2008). In addition, all diffusiontensor information is present in the DTI atlas. This information cantherefore be used to drive the image alignment to this template.

The goal of this study is to examine the effect of the DTI atlasselection on the obtained VBA results. To this end, the VBA sensitivityand specificity to detect WM degeneration after coregistration of alldata sets to different atlases is investigated using simulated DTI datasets (Van Hecke et al., 2009). In these data sets, the size, shape,location, and magnitude of the pathology is known, so they can serveas a ground truth in analysing VBA results (Van Hecke et al., 2010a). Inaddition to the simulations, real DTI data sets of multiple sclerosis(MS) patients and control subjects are examined with VBA usingdifferent DTI templates. These templates are (i) a study-specificpopulation-based template, (ii) a study-specific subject-based atlas inMNI space, and (iii) the ICBM-81 template of Mori et al. (2008).

Methods

Experiment 1: Simulated DTI data sets

In Experiment 1, different VBA analyses were performed. In eachVBA analysis, 20 healthy subject and 20 patient simulated DTI data setswere compared, which were constructed using the method of VanHecke et al. (2009). In this study, the transverse diffusivity (the averageof the second and third eigenvalues (L23) was increased by sevendifferent levels, i.e. by 10%, 20%, 30%, 40%, 50%, 60%, and 70%, to simulateamicrostructural breakdown in 19WMstructures of the 20 patient datasets. The corresponding FA reduction is in agreement with previous DTIfindings in multiple sclerosis patients (Audoin et al., 2007; Cercignaniet al., 2002; Ge et al., 2004; Hasan et al., 2005; Oh et al., 2004; Yu et al.,2007). The location of the simulated WM degeneration is displayed inFig. 1. To summarize, seven different VBA analyses were performed, inwhich 20 healthy data sets were compared to 20 patient data sets withdifferent levels of L23 increase. These sevenVBA analyseswere repeated5 times, in which the estimation of the inter-subject variability and thenoise distribution were varied (Van Hecke et al., 2009).

Experiment 2: Real DTI data sets

Twenty patients with definite MS according to the McDonald criteriawere included in this study (age: 42±9 years; 8 M, 12 F; years ofeducation: 14±2; all right handed) (McDonald et al., 2001). Enrolledsubjects did not have a relapse for at least 30 days before entry into thestudy, did not use sedatives, and had a visual acuity above 20−40, asmeasured on a Snellen chart. The expanded disability status scale (EDSS)

of the patients varied between 0 and 7 (Kurtzke, 1983). Thirteen patientshad a relapse-remitting type, and seven had a secondary progressive typeof MS. A control group of 20 healthy volunteers was matched to the MSpatient group for age, gender and educational level (age: 42±10 years;8 M, 12 F; years of education: 14±2; all right handed) (Van Hecke et al.,2010b). Control subjects usingmedication, having afirst or seconddegreerelative with MS, or having visual impairment were excluded. The studywas approved by the hospital ethics committee and all subjects gavewritten informed consent before entering the study.

DTI data setswere obtained on a1.5 TMR scanner (Siemens, Erlangen,Germany) using an SE-EPI sequence with the following acquisitionparameters: TR: 10.4 s; TE: 100 ms;diffusiongradient strength: 40 mT/m;FOV=256×256mm2;numberof slices=60; voxel size=2×2×2mm3;b-value of 700 s/mm2; acquisition time: 12 min 18 s. Diffusion measure-ments were performed along 60 directions with 10 b0-images. Beforeestimating the diffusion tensor with a non-linear least squares method,the diffusion-weighted images were corrected for subject motion andeddy current induced geometric distortions, including the appropriate b-matrix rotation (Leemans and Jones, 2009).

DTI templates

In this study, the effect of three DTI templates on the VBA results isexamined:

• A group-wise population-specific DTI atlas was constructed asdescribed in Van Hecke et al. (2008), including both the healthy

568 W. Van Hecke et al. / NeuroImage 55 (2011) 566–573

subject and the patient DTI information. In the population-specificatlas approach, non-rigid deformationfields are calculated between alldata sets of the subject group (Van Hecke et al., 2008). Subsequently,each data set is transformed with the mean deformation field to allother data sets. Finally, these transformed images are averaged tocreate the population-specific atlas. The non-rigid coregistrationmethod that was used is based on a viscous fluid model, in whichmutual information is thereby used as a cost function to iterativelycompare the diffusion tensor elements. (Van Hecke et al., 2007). Apopulation-specific DTI atlas was constructed for both experimentsseparately.In the remainder of themanuscript, this population-specificDTI atlas isreferred to as “PA” (population-based atlas).

• Acustomized subject-basedDTI atlas that is situated inMNI space. Fromthe EPI MNI template, a custom FA-based template was constructed asdescribed in Jones et al. (2002) andMori et al. (2008), using 40 healthysubject DTI data sets (age range: 18−50 years) thatwere acquiredwiththe aforementionedMRprotocol. These40data sets,whicharedifferentfrom the 20 healthy subject data sets that were used for the VBA, werenon-rigidly transformed to MNI space to construct the MNI template.During the construction of both the population-specific and the MNIatlas, the same image alignment algorithmswere applied (i.e. affineandtensor element non-rigid coregistration based on a viscous fluid modeland mutual information) in order to eliminate potential confoundingfactors in the analysis. In both atlases, the tensor information is present,which can be used to guide the non-rigid coregistration process. Inaddition, both atlases are averages of the same number of data sets,whichwere acquiredwith the sameMRI protocol, to obtain similar SNRproperties. Note that this viscous fluid based non-rigid coregistrationmethodwas also used in the VBApipeline to align all data to the atlases.In the remainder of the manuscript, this customized subject-based DTIatlas in MNI space is referred to as “SA” (subject-based atlas).In the remainder of this paper, we will refer to this template as the“subject-based MNI atlas.”

• The ICBM-81 DTI atlas that was developed by Mori et al. (2008). Incontrast to the other templates, this atlas is constructed from 81 DTIdata sets, which are acquired with a different acquisition protocolthan the data sets that are coregistered to this atlas in the VBAanalysis. In contrast to the two previous atlases, an affine imagealignment method was used to construct the ICBM-81 DTI atlas.In the remainder of the manuscript, the ICBM-81 DTI atlas of Moriet al. (2008) is referred to as “MA” (Mori's atlas).

VBA pipeline

The following VBA approachwas used to compare the simulated aswell as the real DTI data sets between healthy controls and diseasedsubjects:

• All data sets were coregistered to the atlas with an affine transforma-tion using MIRIT (Multi-modality Image Registration using Informa-tion Theory) based on the FA maps (Maes et al., 1997). Thepreservation of principal direction (PPD) tensor reorientation strategywas incorporated (Alexander et al., 2001; Leemans et al., 2005).

• The affinely aligned data setswere coregistered to the templatewith anon-rigid coregistration method using the viscous fluid model withmutual information as a cost function (Van Hecke et al., 2007). Again,the PPD approach was applied to reorient the tensors after thetransformation, including a smoothingof thefinal deformationfield tocorrect for small errors in the Jacobianmatrix (VanHecke et al., 2007).

• The resulting images were smoothed with an adaptive, anisotropicsmoothing kernel (FWHM=3mm) (Sijbers et al., 1999; Van Heckeet al., 2010a). By using an adaptive, anisotropic smoothing kernel,the WM boundaries are preserved, resulting in a reduction of thepartial volume averaging of WM tissue with gray matter orcerebrospinal fluid, compared to the generally applied isotropic

filter method. This will increase the sensitivity and specificity of thepathology detection in a VBA analysis, as was recently demonstratedon simulated DTI data sets (Van Hecke et al., 2010a).

• The FA and mean diffusivity (MD) values of the healthy subject andthe pathologic data sets were compared in each voxel using a non-parametric Mann-Whitney U test. A correction for multiplecomparisons based on the false discovery rate (FDR) (FDR thresholdof pb0.05) was applied to account for multiple testing (Benjaminiand Hochberg, 1995; Storey, 2003).

Note that, since all parameter settings are identical during the VBApipeline, differences in obtained VBA results can only be attributed tothe atlas selection.

Assessment of image alignment accuracy

In order to evaluate the image alignment accuracy, the coefficientof variance (COV) was calculated by dividing the variance of thediffusion measure in each voxel after image alignment to a templateby the mean of that diffusion measure in the same voxel. A thresholdwas applied to only take the COV in theWM into account (FAN0.2). Inaddition, the cumulative distribution function (CDF) was calculatedfor FA and MD in both atlas spaces.

Results

Experiment 1: Simulated DTI data sets

In order to evaluate the image alignment accuracy, the COV andcorresponding CDF are displayed for both FA and MD in each atlasspaces (Fig. 2). These results demonstrate higher image align-ment accuracy in the “PA” compared to the “SA” and “MA” for bothFA and MD.

In Fig. 3, the VBA results are shown after coregistration of allsimulatedDTI data sets to the “SA,” “MA,” and the “PA” space. The voxelsin which the FA (Fig. 3a) or MD (Fig. 3b) values were significantlydifferent (FDR threshold of pb0.05) between the healthy and patientsimulated data sets are colored inwhite. The L23was thereby increasedby 30% in 19 WM locations that are outlined in Fig. 1 in all patientsimulated data sets. In order to evaluate the accuracy of the VBA results,these results can be compared with the location of the ground truthsimulated pathologies (see Fig. 1). The number of true positive and falsepositive results is displayed for different levels of L23 increase in Fig. 3cand d, respectively. A pathology was assigned “detected”when at leastone significant voxel after the Benjamini-Hochberg correction formultiple comparisons was found in the WM structure in which L23was increased. As visualized in Fig. 3c, more true positive findings areobserved when the L23 increase in the simulated diseased region islarger. In addition, it can be observed that more true positive and lessfalsepositive results are reported in the “PA” space compared to the “SA”as well as the “MA” space.

Experiment 2: Real DTI data sets

In this experiment voxel-based statistical tests were used to comparethe FA andMD of 20 healthy subjects and 20MS patients in the “PA,” the“MA,” and the “SA” space. Similarly as in Fig. 2, theCOVandcorrespondingCDF for MD and FA in the three atlas spaces are shown in Fig. 4.

In Fig. 5, the VBA results of Experiment 2 are displayed aftercoregistration to the “SA,” “MA,” and “PA.” The VBA results clearlydepend on the atlas space to which all data sets are normalized. FAdifferences between the groups were found in the inferior longitu-dinal fasciculus, fornix, forceps major, and corona radiata in the “SA”space and in the inferior longitudinal fasciculus, fornix, splenium,genu, and body of the corpus callosum, forceps minor, forceps major,corona radiata, and cingulum in the “PA” space. FA differences were

Fig. 2. The coefficient of variance (COV) of the FA and MD of the simulated DTI data sets is mapped for each voxel in the “SA” (blue), “MA” (purple), as well as the “PA” (red) atlasspace. In addition, the cumulative distribution function (CDF) of the COV is shown for FA and MD in all atlas spaces.

569W. Van Hecke et al. / NeuroImage 55 (2011) 566–573

observed in the forceps major in the “MA” space. In addition, MDdifferences between groupswere observed in the inferior longitudinalfasciculus, splenium and body of the corpus callosum, forceps major,and corona radiata after alignment to the PA and in the corpuscallosum after coregistration to the “MA.” No MD differences werefound in the “SA” space.

Discussion

Although VBA is a promising post-processing approach to analyzegroups of DTI data sets, VBA results depend on the parameter selectionduring the processing pipeline. For instance, it has already beendemonstrated that the VBA results are determined by the imagealignment procedure aswell as by the smoothingmethod and smoothingkernelwidth (Foonget al., 2002; Park et al., 2004; Jones et al., 2005; Zhanget al., 2007; Sage et al., 2009; Van Hecke et al., 2010a). Aligning DTI datasets of different subjects to a common atlas or template is notstraightforward, due to the local inter-subject variations of the brainmorphology. As a consequence, many papers have highlighted theimportance of image coregistration in a group analysis andmuchwork isdone on developing high-dimensional image alignment algorithms thatare specifically designed for DTI data sets (Bookstein, 2001; Ashburnerand Friston, 2001; Zhang et al., 2007; Sage et al., 2009; Van Hecke et al.,2007). Since the atlas is the reference image to which all data sets arealigned, the template selection can also affect the VBA results:

(I) The magnitude of the deformation fields of all data sets tothe atlas should be minimized to reduce residual spatialmisalignment after non-rigid coregistration. By using an MNI-based template as the reference space in a VBA analysis, thedeformation fields of the subject group DTI data sets to this atlascan be large, since the MNI template was constructed using dataof relatively young healthy subjects. Especially when dealing

with patient groups or very young or old subjects, the alignmentto a standard atlas can result in unbalanced coregistrationinaccuracies and false positive as well as false negative VBAresults.

(II) The image information in the atlas space determines theinformation that can be used to estimate the non-rigiddeformation fields. Since it has been demonstrated by severalresearchers that the use of tensor information significantlyimproves the accuracy of the image alignment, this informationneeds to be present in the DTI atlas. Until recently, with theintroduction of the ICBM-81 atlas in MNI space by Mori et al.(2008), only anatomical MR information was present in thisMNI template. As a consequence, in most studies, the imagealignment of the DTI data sets to this MNI template was basedon this anatomical MR information, resulting in residual imagealignment inaccuracies of the WM structures (Sage et al.,2009).

(III) It is generally accepted that affine transformations cannot dealwith the local morphological discrepancies between differentsubjects. In order to minimize local differences in brain shapeacross subjects, non-rigid coregistration algorithms are re-quired. In this context, it is important to note that affine atlasesare not optimal templates for the non-rigid transformation ofdata sets.

To address these issues, we recently proposed the use of a non-rigid population-specific atlas that contains all tensor information(Van Hecke et al., 2008). Since the atlas is constructed from the study-specific group that is analyzed, the deformation fields of all DTI datasets of the study group to this population-specific atlas are minimized.Furthermore, a high-dimensional viscous fluid model that included alltensor information was used to align all images to the atlases, whichhas been shown to yield reliable DTI image alignment (Van Hecke

Fig. 3. In (a) and (b), the VBA results are shown after coregistration of the simulated data sets to the “SA” (blue), “MA” (purple) and the “PA” (red) atlas space for FA and MD,respectively. In (c) and (d), the number of true positive and false positive results are depicted for different levels of L23 increase in the locations of simulated white matterdegeneration that are visualized in Fig. 1.

570 W. Van Hecke et al. / NeuroImage 55 (2011) 566–573

et al., 2007). It should be noted, however that although we used thetensor elements as input information for our non-rigid coregistrationmethod, recent work suggests that a high image alignment accuracycan also be obtained when structural MR information is used in acombined structural and surface algorithm (Zöllei et al., 2010).

In this study, we evaluated the effect of the template selection onthe VBA results, by aligning both simulated DTI data sets and real DTIdata sets of control subjects and MS patients to a“SA,” “MA,” and “PA.”It is important to note that both the “SA” and “PA” were constructedfrom an identical number of data sets, which were acquired with thesame protocol, using the same affine and non-rigid coregistrationalgorithms. Since these factors are the same for both analyses, webelieve that the observed improvement in coregistration accuracy canbe assigned to the selection of the template. In addition, as all post-processing options of the VBA pipeline were identical for allexperiments (i.e. the same coregistration algorithm, smoothing

method, smoothing kernel, statistical tests, and post-hoc correction),variations of the VBA results are only related to the template selectionand are not biased by other VBA related factors.

We demonstrated a clear effect of template selection on the VBAresults, bothwhen comparing simulated DTI data and when comparingreal DTI data sets. Our results on simulated data sets suggest a higherVBA sensitivity and specificity to detect the simulated pathologies afteraligning the data sets to the “PA” (see Fig. 3). As shown in Fig. 2, this canbe explained by the fact that higher coregistration accuracy wasachieved after aligning all data sets to the “PA” compared to the “SA” or“MA.” Similar results were obtained when comparing real DTI data setsof healthy subjects andMS patients (see Fig. 5). FA differences betweenboth groupswere found in a larger number ofWM locations in the “PA”space compared to the “SA” space, with an FA decrease in MS patientscompared to controls in 9, 4, and 1WM locations in the “PA,” “SA,” and“MA” space, respectively. The results for MD were more striking, as we

Fig. 4. The coefficient of variance (COV) of the FA andMD ismapped for each voxel in the “SA” (blue), “MA” (purple), aswell as the “PA” (red) atlas space after coregistration of the real DTIdata sets of healthy subjects and multiple sclerosis patients. In addition, the cumulative distribution function (CDF) of the COV is shown for FA and MD in all three atlas spaces.

571W. Van Hecke et al. / NeuroImage 55 (2011) 566–573

found an increase of MD in MS patients compared to controls in 5 WMstructures in the “PA” space, and only in the corpus callosum in the“MA,”whereas noMDdifferenceswere observed in the “SA” atlas. Thesediscrepancies across the VBA results in the different atlas spaces canagain be explained by differences in the image alignment accuracy, ascan be observed in Fig. 4. Note that, in general, the COV of FAwas highercompared to the COV ofMD, due to the higher FA contrast betweenWMand GM. On the other hand, a high MD contrast exists between CSF andsurrounding tissue, which explains the higher COV of MD around theventricles.

As shownby the results in Figs. 2–5, both sensitivity and specificity ofthe VBA results deterioratedwhen the “MA”was used, compared to the“PA” and “SA.” At first instance, this outcome may be surprising, given

Fig. 5. The VBA results that compare FA and MD between healthy subjects and multiple scl

the higher SNR of the “MA” compared to both “SA” and “PA.” There stillare, however, several arguments why a better coregistration perfor-mance – and subsequently,more reliable VBA results–maybe expectedif the “SA” and, especially, “PA” are used as the template compared to the“MA.” First, the “MA” is based on an affine transformation model,whereas the “SA” and “PA” are based on a high-dimensional non-affinedeformationmodel (preceded by an affine registration step). As a result,fine details in the “MA” are blurred more heavily compared to the “SA”and “PA.” This, in turn, affects both the precision and accuracy withwhich the individual data sets can be aligned. Our results indeed showsmaller residual misalignments between the individually registereddata sets when the “SA” and “PA” are used as a reference template.Second, It is well known that DTI data suffer from non-linear geometric

erosis patients in the “SA” (blue), “MA” (purple), as well as the “PA” (red) atlas space.

572 W. Van Hecke et al. / NeuroImage 55 (2011) 566–573

distortions, which are highly scanner-dependent (e.g. induced by theinhomogeneity of the static B0 field/different performance in shim-ming/etc.): at higher magnetic field strengths, for instance, thesedistortions become more prominent. So in principal, registering datasets obtained with that same scanner would be more accurate thanregistering data sets (with the same SNR) acquired with a differentscanner, since the deformation fields between the data from differentscanners would be (on average, at least) larger than those between thedata from the samescanner. Since the “MA” is constructedwithdata setsfrom a different scanner than the data sets used in this work forcoregistration to the templates, the observed difference in our VBAresults between the “MA” and the other atlases (“SA” and “PA”) may(partly) be caused by this principle. Third, as aforementioned, thesubjects that were used to construct the “MA” (all healthy middle-agedadolescents)maynot represent the subject group under investigation inan appropriate (unbiased) way. Such a difference in characteristicsbetween those subjects thatwereused to build the atlas and the subjectsthat are included in a VBA study may also increase the deformationfields needed to align the subjects. Finally, it is known that registrationaccuracy may depend on the initialization: if larger transformations areneeded to align individual data from native space to template space, thecoregistration quality is typically worse. This ‘geometric distance’between individual data sets and the atlas is the lowest (by constructionof the group-wise approach) for the “PA” and, in general, highest for the“MA” (usually higher than “SA” depending on the FOV). Again, our VBAresults may be (partly) explained by this coregistration property.

Although, the use of a population-specific atlas improves thereliability of VBA results, there are some limitations to this approach.The widely used standard MNI atlas, in particular, has some distinctadvantages compared to population-specific templates. First, since theMNI atlas contains anatomic and cytoarchitectonic labels in a standardcoordinate space, it is easier to report and compare results (and specificcoordinates) across studies. Studies using population-specific atlaseswould rather report the WM bundles/structures in which diffusionchanges were observed. An alternative approach could be to use apopulation-specific DTI atlas and subsequently transform the statisticalmap to MNI space. Second, as atlases in MNI space are generallyconstructed froma largenumber of data sets (i.e.,more than thenumberof subjects that is commonly included in a VBA analysis), the SNR of theresulting MNI atlas is very high. Especially when VBA studies areperformed with a relatively low number of data sets, the use of an MNIatlas might be advantageous, since the resulting population-specificatlas will have a lower SNR. However, the exact relationship betweenthe SNR of a specific atlas and the coregistration accuracy of data sets tothis atlas is still unclear and is an interesting topic for futurework. Third,inMNI space, other image information (suchas anatomical or functionalMRI information) is generally available, which can be convenient formulti-modal studies. However, by using a population-specific DTI atlasand subsequently transforming the statistical map to MNI space, thispotential disadvantage of population-specific atlases may again becircumvented. Finally, when (larger) lesions are present in some of theparticipants, the population-specific atlas may be biased towards thesesubjects. It remains unclear to which extent the image alignmentaccuracy to the MNI space would be affected in this case. In general,when many subjects with lesions are present in the group study, it isprobably better to use ROI-based analyses instead of VBA.

In conclusion, we have demonstrated that the selection of thetemplate space can affect the VBA results of DTI data sets. Our resultson simulated and real DTI data sets suggest that the VBA results aremore reliable after coregistration of the images to a population-specific, non-rigid DTI atlas.

References

Alexander, D., Pierpaoli, C., Basser, P., Gee, J., 2001. Spatial transformations of diffusiontensor magnetic resonance images. IEEE Trans. Med. Imaging 20 (11), 1131–1139.

Ashburner, J., Friston, K., 2000. Voxel-based morphometry — the methods. Neuroimage11 (6), 805–821.

Ashburner, J., Friston, K., 2001. Why voxel-based morphometry should be used.Neuroimage 14 (6), 1238–1243.

Audoin, B., Guye, M., Reuter, F., Au Duong, M.V., Confort-Gouny, S., Malikova, I.,Soulier, E., Viout, P., Chrif, A.A., Cozzone, P.J., Pelletier, J., Ranjeva, J.P., 2007.Structure of WM bundles constituting the working memory system in earlymultiple sclerosis: a quantitative DTI tractography study. Neuroimage 36 (4),1324–1330.

Benjamini, Y., Hochberg, Y., 1995. Controlling the false discovery rate: a practical andpowerful approach to multiple testing. J. Roy. Statist. Soc. Ser. B 57, 289–300.

Bookstein, F.L., 2001. Voxel-based morphometry should not be used with imperfectlyregistered images. Neuroimage 14 (6), 1454–1462.

Cercignani, M., Bozalli, M., Iannucci, G., Comi, G., Filippi, M., 2002. Intra-voxel and inter-voxel coherence in patients with multiple sclerosis assessed using diffusion tensorMRI. J. Neurol. 249 (7), 875–883.

Douaud, G., Smith, S., Jenkinson, M., Behrens, T., Johansen-Berg, H., Vickers, J., James, S.,Voets, N., Watkins, K., Matthews, P.M., James, A., 2007. Anatomically related greyand white matter abnormalities in adolescent-onset schizophrenia. Brain 130 (9),2375–2386.

Foong, J., Symms, M.R., Barker, G.J., Maier, M., Miller, D.H., Ron, M.A., 2002. Investigatingregional white matter in schizophrenia using diffusion tensor imaging. NeuroRe-port 13 (3), 333–336.

Ge, Y., Law, M., Johnson, G., Herbert, J., Babb, J.S., Mannon, L.J., Grossman, R.I., 2004.Preferential occult injury of corpus callosum in multiple sclerosis measured bydiffusion tensor imaging. J. Magn. Reson. Imaging 20 (1), 1–7.

Hasan, K.M., Gupta, R.K., Santos, R.M., Wolinsky, J.S., Narayana, P.A., 2005. Fractionaldiffusion tensor anisotropy of the seven segments of the normal-appearing whitematter of the corpus callosum in healthy adults and relapsing remitting multiplesclerosis. J. Magn. Reson. Imaging 21 (6), 735–743.

Hsu, J.L., Leemans, A., Bai, C.H., Lee, C.H., Tsai, Y.F., Chiu, H.C., Chen, W.H., 2008. Genderdifferences and age-related white matter changes of the human brain: a diffusiontensor imaging study. Neuroimage 39 (2), 566–577.

Hsu, J.L., Van Hecke,W., Bai, C.H., Lee, C.H., Tsai, Y.F., Chiu, H.C., Jaw, F.S., Hsu, C.Y., Leu, J.G.,Chen,W.H., Leemans, A., 2010.Microstructuralwhitematter changes innormal aging:a diffusion tensor imaging study with higher-order polynomial regression models.Neuroimage 49 (1), 32–43.

Jones, D., Griffin, L., Alexander, D., Catani, M., Horsfield, M., Howard, R., Williams, S.,2002. Spatial normalization and averaging of diffusion tensor MRI data sets.Neuroimage 17 (2), 592–617.

Jones, D., Symms, M., Cercignani, M., Howarde, R., 2005. The effect of filter size on VBManalyses of DT-MRI data. Neuroimage 26 (2), 546–554.

Kurtzke, J.F., 1983. Rating neurologic impairment in multiple sclerosis: an expandeddisability status scale (EDSS). Neurology 33, 1444–1452.

Kyriakopoulos, M., Vyas, N.S., Barker, G.J., Chitnis, X.A., Frangou, S., 2007. A diffusion tensorimagingstudyofwhitematter inearly-onset schizophrenia. Biol Psychiatryiatry 63 (5),519–523.

Leemans, A., Jones, D.K., 2009. The B-matrix must be rotated when correcting forsubject motion in DTI data. Magn. Reson. Med. 62 (1), 1336–1349.

Leemans, A., Sijbers, J., De Backer, S., Vandervliet, E., Parizel, P., 2005. Affinecoregistration of diffusion tensor magnetic resonance images using mutualinformation. Lecture Notes in Computer Science, pp. 523–530.

Maes, F., Collignon, A., Vandermeulen, D., Marchal, G., Suetens, P., 1997. Multimodalityimage registration by maximization of mutual information. IEEE Trans. Med. Imaging16 (2), 187–198.

McDonald, W.I., Compston, A., Edan, G., Goodkin, D., Hartung, H.P., Lublin, F.D.,McFarland, H.F., Paty, D.W., Polman, C.H., Reingold, S.C., Sandberg-Wollheim, M.,Sibley, W., Thompson, A., van den Noort, S., Weinshenker, B.Y., Wolinsky, J.S.,2001. Recommended diagnostic criteria for multiple sclero- sis: guidelines fromthe International Panel on the diagnosis of multiple sclerosis. Ann. Neurol. 50,121–127.

Mori, S., Oishi, K., Jiang, H.Y., Jiang, L., Li, X., Akhter, K., Hua, K.G., Faria, A.V., Mahmood,A., R.W., 2008. Stereotaxic white matter atlas based on diffusion tensor imaging inan ICBM template. Neuroimage 40 (2), 570–582.

Mori, S., Oishi, K., Faria, A.V., 2009. White matter atlases based on diffusion tensorImaging. Curr. Opin. Neurol. 22 (4), 362–369.

Oh, J., Henry, R.G., Genain, C., Nelson, S.J., Pelletier, D., 2004. Mechanisms of normalappearing corpus callosum injury related to pericallosal T1 lesions in multiplesclerosis using directional diffusion tensor and 1 H MRS imaging. J. Neurol.Neurosurg. Psychiatry 75 (9), 1281–1286.

Park, H.-J., Kubicki, M., Shenton, M., Guimond, A., McCarley, R., Maier, S., Kikinis, R.,Jolesz, F., Westin, C.-F., 2003. Spatial normalization of diffusion tensor MRI usingmultiple channels. Neuroimage 20 (4), 1995–2009.

Park, H.-J., Westin, C.-F., Kubicki, M., Maier, S., Niznikiewicz, M., Baer, A., Frumin, M.,Kikinis, R., Jolesz, F., McCarley, R., Shenton, M., 2004. White matter hemisphereasymmetries in healthy subjects and in schizophrenia: a diffusion tensorMRI study.Neuroimage 23 (1), 213–223.

Sage, C., Van Hecke, W., Peeters, R., Robberecht, W., Sijbers, J., Leemans, A., Parizel, P.M.,Marchal, G., Sunaert, S., 2009. Quantitative diffusion tensor imaging in amyotrophiclateral sclerosis: revisited. Hum Brain Map 30 (11), 3657–3675.

Sijbers, J., den Dekker, A., Verhoye, M., Van der Linden, A., Van Dyck, D., 1999. Adaptiveanisotropic noise filtering magnitude MR data 17 (10), 1533–1539.

Smith, S.M., Jenkinson, M., Johansen-Berg, H., Rueckert, D., Nichols, T.E., Mackay, C.E.,Watkins, K.E., Ciccarelli, O., Cader, M.Z., Matthews, P.M., Behrens, T.E., 2006. Tract-based spatial statistics: voxelwise analysis of multi-subject diffusion data. Neuro-image 4 (31), 1487–1505.

573W. Van Hecke et al. / NeuroImage 55 (2011) 566–573

Storey, J.D., 2003. The positive false discovery rate: a Bayesian interpretation and the q-value. Ann. Stat. 31, 2013–2035.

Sundgren, P.C., Dong, Q., Gomez-Hassan, D., Mukherji, S.K., Welsh, R., 2004. Diffusiontensor imaging of the brain: review of clinical applications. Neuroradiology 46 (5),339–350.

Van Hecke, W., Leemans, A., D'Agostino, E., De Backer, S., Vandervliet, E., Parizel, P.M.,Sijbers, J., 2007. Nonrigid coregistration of diffusion tensor images using a viscousfluid model and mutual information. IEEE Trans. Med. Imaging 26 (11), 1598–1612.

VanHecke,W., Leemans, A., D'Agostino, E.,Maes, F., De Backer, S., Vandervliet, E., Sijbers, J.,Parizel, P.M., 2008. On the construction of an inter-subject diffusion tensor magneticresonance atlas of the healthy human brain. Neuroimage 43 (1), 69–80.

Van Hecke, W., Sijbers, J., De Backer, S., Poot, D., Parizel, P.M., Leemans, A., 2009. On theconstruction of a ground truth framework for evaluating voxel-based diffusiontensor MRI analysis methods. Neuroimage 46 (3), 692–707.

Van Hecke, W., Leemans, A., De Backer, S., Jeurissen, B., Parizel, P.M., Sijbers, J., 2010a.Comparing isotropic and anisotropic filtering for voxel based DTI analyses: asimulation study. Hum. Brain Mapp. 31 (1), 98–114.

Van Hecke, W., Nagels, G., Leemans, A., Vandervliet, E., Sijbers, J., Parizel, P.M., 2010b.Correlation of white matter damage and cognitive dysfunction in patients withmild and moderate multiple sclerosis. J Magn Reson Imag 31 (6), 1492–1498.

White, T., Kendi, A.T.K., Lehricy, S., Kendi, M., Karatekin, C., Guimaraes, A., Davenport, N.,Schulz, S.C., Lim, K.O., 2007. Disruption of hippocampal connectivity in children andadolescents with schizophrenia - a voxel-based diffusion tensor imaging study.Schizophr. Res. 90 (1–3), 302–307.

Yu, C.S., Zhu, C.Z., Li, K.C., Xuan, Y., Qin, W., Sun, H., Chan, P., 2007. Relapsingneuromyelitis optica and relapsing-remitting multiple sclerosis: differentiation atdiffusion-tensor MR imaging of corpus callosum. Radiology 244 (1), 249–256.

Zhang, H., Avants, B.B., Yushkevich, P.A., Woo, J.H., Wang, S., McCluskey, L.F., Elman, L.B.,Melhem, E.R., Gee, J.C., 2007. High-dimensional spatial normalization of diffusiontensor images improves the detection of white matter differences: An example studyusing amyotrophic lateral sclerosis. IEEE Trans. Med. Imaging 26 (11), 1585–1597.

Zöllei, L., Stevens, A., Huber, K., Kakunoori, S., Fischl, B., 2010. Improved tractographyalignment using combined volumetric and surface registration. Neuroimage 51 (1),206–213.