www.elsevier.com/locate/ynimg

NeuroImage 38 (2007) 114–123

Brain tissue segmentation based on DTI data

Tianming Liu,a,c Hai Li,a,b,c Kelvin Wong,a,c Ashley Tarokh,d

Lei Guo,b and Stephen T.C. Wonga,c,⁎

aDepartment of Radiology, The Methodist Hospital, Houston, TX, USAbSchool of Automation, Northwestern Polytechnic University, Xi’an, ChinacCornell Weill Medical College, USAdFunctional and Molecular Imaging Center, Department of Radiology, Brigham and Women’s Hospital, Boston, MA, USA

Received 4 April 2007; revised 2 June 2007; accepted 4 July 2007Available online 13 July 2007

We present a method for automated brain tissue segmentation based onthe multi-channel fusion of diffusion tensor imaging (DTI) data. Themethod is motivated by the evidence that independent tissuesegmentation based on DTI parametric images provides complemen-tary information of tissue contrast to the tissue segmentation based onstructural MRI data. This has important applications in definingaccurate tissue maps when fusing structural data with diffusion data.In the absence of structural data, tissue segmentation based on DTIdata provides an alternative means to obtain brain tissue segmentation.Our approach to the tissue segmentation based on DTI data is toclassify the brain into two compartments by utilizing the tissue contrastexisting in a single channel. Specifically, because the apparent diffusioncoefficient (ADC) values in the cerebrospinal fluid (CSF) are morethan twice that of gray matter (GM) and white matter (WM), we useADC images to distinguish CSF and non-CSF tissues. Additionally,fractional anisotropy (FA) images are used to separate WM from non-WM tissues, as highly directional white matter structures have muchlarger fractional anisotropy values. Moreover, other channels toseparate tissue are explored, such as eigenvalues of the tensor, relativeanisotropy (RA), and volume ratio (VR). We developed an approachbased on the Simultaneous Truth and Performance Level Estimation(STAPLE) algorithm that combines these two-class maps to obtain acomplete tissue segmentation map of CSF, GM, and WM. Evaluationsare provided to demonstrate the performance of our approach.Experimental results of applying this approach to brain tissuesegmentation and deformable registration of DTI data and spoiledgradient-echo (SPGR) data are also provided.© 2007 Elsevier Inc. All rights reserved.

Introduction

Brain tissue segmentation has important applications in study-ing the structure and function of the brain. A number of methods

⁎ Corresponding author. Department of Radiology, The MethodistHospital Research Institute, Houston, TX, USA.

E-mail address: wong@crystal.harvard.edu (S.T.C. Wong).Available online on ScienceDirect (www.sciencedirect.com).

1053-8119/$ - see front matter © 2007 Elsevier Inc. All rights reserved.doi:10.1016/j.neuroimage.2007.07.002

based on structural MRI data have been proposed for the seg-mentation problem (Zhang et al., 2001; Wells et al., 1996; Phamand Prince, 1999; Dale et al., 1999). In this work, we propose arobust method for automated brain tissue segmentation based onthe multiple-channel fusion in DTI space. Our method can beemployed to define accurate tissue maps when dealing with fusedstructural and diffusion data. This enables us to study the graymatter diffusivity in neurodegenerative and neurological diseases(Liu et al., 2005, 2006). When fusing structural and diffusioninformation, the imperfect alignment of structural MRI data, e.g.,SPGR image, with DTI data results in the problem of hetero-geneous voxels when the anatomic information in the structuraldata is applied to the DTI data. Under the problem of hetero-geneous voxels, the measurements of the GM diffusivity based onthe anatomic information in the SPGR image may fail to reveal thereal diffusion in the GM. Figs. 3h and 3i in Liu et al. (2006)illustrate examples of such a problem. Specifically, following non-rigid co-registration using the UCLA AIR tools (Woods et al.,1998), the GM boundaries of SPGR image are crossing CSF ofADC image. Consequently, the GM voxels in the SPGR imagecorrespond to CSF voxels in the ADC image. Such a problem canoccur for a variety of reasons, including geometric distortion inDTI imaging (Jezzard and Balaban, 1995), partial volume effect(Helenius et al., 2002; Gonza’lez Ballester et al., 2002), reslicingand interpolation of DTI data, and errors in co-registration.

Recently, we proposed a two-channel fusion method to removeheterogeneous voxels (Liu et al., 2005, 2006). In that work, weperformed tissue segmentations in both SPGR space (Zhang et al.,2001) and DTI space, and combined the results to obtain the mostconservative definition of GM tissue, which is the consensus ofboth spaces. For example, Fig. 6 in Liu et al. (2006) shows theconservative definition of GM after heterogeneous voxels removalis performed in Fig. 3 of Liu et al., (2006). In order to performtissue segmentation in the DTI space, the ADC image was used todistinguish the CSF and non-CSF. The technique takes advantageof the fact that the ADC values in CSF are more than twice as highas the GM and WM values. Meanwhile, the FA image was used toseparate the WM from the non-WM tissues, as highly directional

115T. Liu et al. / NeuroImage 38 (2007) 114–123

white matter structures have much larger fractional anisotropyvalues.

Our prior approaches employ only ADC and FA channels toclassify brain tissues (Liu et al., 2005, 2006), and in this work, ourmethod of tissue segmentation in DTI space is further improved byusing the following seven individual channels: ADC, eigenvalues(λ1, λ2, λ3), FA, RA, and VR. The Simultaneous Truth andPerformance Level Estimation (STAPLE) algorithm (Warfield et al.,2004) is enlisted to combine these two-class maps and obtain acomplete segmentation map of CSF, GM, and WM. Extensiveevaluations and comparison studies are provided to demonstrate thereasonably good performance of our improved method. Expe-rimental results of applying the proposed method to brain tissuesegmentation and deformable registration of DTI data and SPGRdata are also presented.

Method

Background

Diffusion-weighted imaging (DWI) and DTI permit in vivomeasures of the diffusion of water molecules in living tissues (LeBihan, 1991). The diffusion of water molecules is typically repre-sented by unrestricted Brownian motion, a particular tissue structurecan preferentially restrict the molecular motion, which leads toanisotropic diffusion that is measured by DWI and DTI (Le Bihan,1991; Bammer, 2003). As an approximation, the measured diffusioncan be modeled as an anisotropic Gaussian, parameterized by thediffusion tensor in each voxel (Basser et al., 1994), in order to createa 3-D field of diffusion tensors. Diffusion tensor measurementsprovide a rich dataset from which a measure of diffusion anisotropycan be obtained in various ways through the application ofmathematical formulas and recalculation of the underlying eigen-values (Bammer, 2003; Moseley et al., 1990; Le Bihan et al., 2001;

Fig. 1. Illustration of the computational framework of tissue segmentation based osegmentation; (3) WM/non-WM segmentation; (4) multi-channel fusion to obtain C(6) combining the CSF/non-CSF and WM/non-WM maps into a complete CSF/W

Basser and Jones, 2002). The ADC, λ1, λ2, and λ3 channels arewell suited for the measuring of overall diffusivity, whereasanisotropy can be represented by FA, RA, or VR (Sundgren et al.,2004). In Sundgren et al., 2004, it was reported that the diffusivityvalues of the CSF are more than double the GM and WM values.This is because water diffusion in CSF is much less restricted thanthose in the GM and WM tissues (Johanna et al., 2002). Therefore,we can use these four channels of images to segment CSF from non-CSF tissues. Also, because highly directional white matter structureshave much larger fractional anisotropy values, the FA, RA, and VRimages can be used to separate WM from non-WM tissues.Moreover, since all these images intrinsically share the same DTIspace, the CSF/non-CSF and WM/non-WM images can becombined into a complete CSF/GM/WM segmentation map in theDTI space without the need of any registration.

Overview

Our computational framework for tissue segmentation based onDTI data consists of six steps, as summarized in Fig. 1. The firststep consists of pre-processing to perform eddy current correctionthrough the use of FSL FDT tools (http://www.fmrib.ox.ac.uk/fsl/fdt/index.html), tensor calculation, and channel image generationusing DTIStudio (http://cmrm.med.jhmi.edu/DTIuser/DTIuser.asp). As a result, seven channels are obtained: ADC, λ1, λ2, λ3,FA, RA, and VR. To reduce noise, we smooth these seven channelsthrough an edge preserving anisotropic diffusion filter (Perona andMalik, 1990). The second step consists of segmenting the braininto CSF and non-CSF compartments by utilizing the tissuecontrasts existing in the first four channels, i.e., ADC, λ1, λ2, λ3.The third step consists of segmenting the brain into WM and non-WM compartments using the last three channels, i.e., FA, RA, andVR, separately. In order to generate the final CSF and non-CSFmap, the fourth step employs the STAPLE algorithm (Warfield et

n DWI/DTI data. There are six steps: (1) Pre-processing; (2) CSF/non-CSFSF/non-CSF map; (5) multi-channel fusion to obtainWM/non-WMmap; andM/GM tissue segmentation.

116 T. Liu et al. / NeuroImage 38 (2007) 114–123

al., 2004) to fuse the tissue segmentation results of the first fourchannels (ADC, λ1, λ2, λ3). The WM and non-WM map isgenerated in the fifth step. The final step consists of combiningboth CSF/non-CSF and WM/non-WM maps to obtain a completetissue map detailing the CSF, WM, and GM segments.

Tissue segmentation based on DTI data

HMRF-EM tissue segmentationIn each individual channel outlined above, the brain is classified

into two classes: either CSF and non-CSF, or WM and non-WM,depending on the channel. The Expectation-Maximization (EM)algorithm, in combination with a Hidden Markov Random Field(HMRF) model (Li, 2001), is used for the two-class tissuesegmentation (Zhang et al., 2001; Liu et al., 2005, 2006). The EMmodel fitting and the MRF Iterated Conditional Modes (ICM)labeling in the HMRF-EM segmentation both require the selectionof an initial parameter set (Zhang et al., 2001). In the literature, the k-means clustering has been widely used for automated selection ofinitial centroids (Zhang et al., 2001; Pham and Prince, 1999).We usethis method for the initial estimation.

Multi-channel fusion

Motivation. The application of the HMRF-EM segmentationmethod to different channels, such as ADC and eigenvalues, resultsin a discrepancy in the segmentation results. Fig. 2 provides anexample where the distribution of the ADC values in the entirebrain along with the segmented CSF of ADC channel and the λ3channel are presented. Fig. 2 indicates that there are no clearboundaries between different tissues in the ADC distribution of thewhole brain (shown in red). Also, the CSF segmentation results ofADC channel (shown in green) and λ3 channel (shown in blue) arequite different. From this, we postulate that segmentations frommultiple channels provide richer information, and an optimalcombination of these results will be more desirable.

Due to the difficulty of obtaining or estimating a known truesegmentation for real data, the performance of the segmentationfrom any given channel is difficult to quantify, and each channel

Fig. 2. The distribution of the ADC values in the whole brain and in the CSFof single-channel segmentation result using ADC and λ3 channel separately.ADC scale is 10−3 mm2/s.

could not be assumed to contribute equally to the combinedsegmentation result. To deal with this problem, we make use of theExpectation-Maximization algorithm for Simultaneous Truth andPerformance Level Estimation (STAPLE) algorithm proposedin Warfield et al. (2004). The algorithm considers a collection ofsegmentations and for each segmentation computes a probabilisticestimate of the true segmentation and a measure of the performancelevel represented by that segmentation. It then proceeds to estimatethe optimal combination of the segmentations in order to obtain aprobabilistic estimate of the true segmentation. This is done byweighting each segmentation according to its estimated perfor-mance level while employing a prior model for the spatial dis-tribution of structures being segmented and spatial homogeneityconstraints. Because we consider seven two-class tissue segmenta-tion maps, the goal of our multi-channel fusion is to combine theseven binary segmentation maps to construct an improved overallsegmentation map and to characterize the segmentation perfor-mance level of every channel, simultaneously.

Fusion of CSF/non-CSF segmentations using STAPLE. Asoutlined above, there are four binary segmentation maps (CSF/non-CSF) associated with four channels (ADC, λ1, λ2, λ3). Let Xbe the random field defined in the whole volume of N voxels, andxi denote the configuration of each voxel i. We have

X ¼ fxi ¼ ðxi1; xi2; xi3; xi4Þjxijaf0; 1g; i ¼ 1; L;N ; j ¼ 1; L 4gð1Þ

where 0 and 1 represent the CSF and non-CSF two tissue types. LetY be the true segmentation map. The segmentation performance ofevery channel is characterized by sensitivity and specificity. Sensiti-vity is the relative frequency of xij=1 when Yi= 1. We define p=(p1, p2, p3, p4)

T as a column vector, whose ith element is thesensitivity parameter for the segmentation on the ith channel.Similarly, specificity is the relative frequency of xij=0 when Yi=0for which we define the column vector of specificity values q=(q1,q2, q3, q4)

T for the four channels considered (Warfield et al., 2004).We then apply the STAPLE algorithm to estimate the true map

via the maximum a posteriori (MAP) method. Specifically, we seeka map Y*, which is an estimate of the true map Y, according to theMAP criterion:

Y* ¼ arg maxYaX

ðf ðX jY ; p; qÞf ðY ÞÞ¼ arg max

YaXðj

i½jjf ðXijjYi; pj; qjÞ�f ðYiÞÞ ð2Þ

where f (Yi) is the prior probability of Yi, and a voxelwise inde-pendence assumption has been made here. Next, assuming that thesegmentation maps are mutually independent, the performancelevel parameter, which maximizes the complete data (X,Y) loglikelihood function, is given by:

ðpj; qjÞ ¼ arg maxpj;qj

ln f ðXj; Y jpj; qjÞ ð3Þ

To estimate the solution of Eq. (3), an EM algorithm is used. Theiteration of the EM algorithm will be performed as:

pðkÞj ¼P

i:Xij¼1Wðk�1ÞiP

iWðk�1Þi

ð4Þ

qðkÞj ¼P

i:Xij¼0ð1�W ðk�1Þi ÞP

ið1�W ðk�1Þi Þ

ð5Þ

117T. Liu et al. / NeuroImage 38 (2007) 114–123

In Eqs. (4) and (5), Wi(k−1) indicates the probability of the true

segmentation at voxel i being equal to 1:

W ðk�1Þi uf Yi ¼ 1jXi; p

ðk�1Þ; qðk�1Þ� �

¼ mðk�1Þi

mðk�1Þi þ nðk�1Þ

i

ð6Þ

where mðk�1Þi ¼ f ðYi¼1Þjj:Xij¼1p

ðk�1Þj jj:Xij¼0ð1� pðk�1Þ

j Þ, nðk�1Þi ¼

f ðYi ¼ 0Þjj:Xij¼0qðk�1Þj jj:Xij¼1ð1� qðk�1Þ

j ÞÞ. Because the true seg-mentation is a binary random variable, the posterior probabilityf ðYi ¼ 0jXi; pðk�1Þ; qðk�1ÞÞ is equal to 1−Wi

(k−1). According to Eq.(2), the fusion model can be set as:

Yi ¼ 0 if Wi b 0:51 if Wiz 0:5

�ð7Þ

In summary, there are two steps in STAPLE for multiple-channelfusion. At each iteration, the first step entails the estimation of theconditional probability of the true segmentation given the segment-ation maps of four channels and previous performance parameterestimates. In the second step, estimation values of the performanceparameters are updated. More details concerning implementation ofthe iterative algorithm can be found in Warfield et al. (2004). Thefusion result is finally obtained through Eq. (7).

Fusion of WM/non-WM segmentations. First, the three channelsof FA, RA, and VR images are separately segmented into WM/non-WM tissue maps using the HMRF-EM method described in theHMRF-EM tissue segmentation subsection. Then, the three tissuesegmentation maps are fused into a complete WM/non-WM tissuemap by using the STAPLE algorithm described in the Fusion ofCSF/non-CSF segmentations using STAPLE subsection. As a result,the FA, RA, and VR channels provide an independent segmentationof the brain tissues into the WM and non-WM compartments.

Experimental results

Datasets

We have applied our tissue segmentation and multiple-channelfusion method to 10 SPGR and DTI datasets. For the SPGR imaging

Table 1Multiple-channel fusion result

SPGR CSF (%) GM (%)

19.7 41.0

DWI/DTI CSF (%) Non-CSF (

⁎ ADC 11.7 88.3λ1 18.2 81.8λ2 15.5 84.5λ3 22.2 77.8

# Y 21.4 78.6

DWI/DTI WM (%) Non-WM (%)

⁎ FA 67.6 32.4RA 62.4 37.6VR 58.6 41.4

# Y 49.5 50.5

“⁎”marks the single-channel segmentation result, and “#”marks the fusion result. Tin SPGR space or DWI/DTI space. The fifth column shows the volume overlap (Ochannel against the corresponding tissue type in SPGR image. The sixth column shovolume agreement). The seventh column shows the segmentation performance lev

settings, a 1.5-Tesla GE Echospeed system was used with coronalseries of contiguous spoiled gradient images. The voxel dimensionsin this case are equal to: 0.9375×0.9375×1.5 mm. For the DTIsettings, a 1.5-Tesla GE Echospeed system was used with the fol-lowing settings. Sequence: maximum gradient amplitudes: 40 mT/M; rectangular FOV: 220×165 mm; 4 mm slice thickness; 1 mminterslice distance; TE: 70 ms; TR: 2500 ms; b value: 1000 s/mm2.DWI images along six non-collinear directions were collected. Moredetails about our imaging settings can be found in Kubicki et al.(2005).

The SPGR data and DTI datasets have been processed using thepre-processing methods as described in Liu et al. (2006). In par-ticular, the SPGR image has been co-registered with the DTIimages using the non-rigid registration method in the UCLA AIRpackage (Woods et al., 1998).

Experiment 1

In the fusion of CSF/non-CSF segmentation maps, the initialvalues (pj

(0), qj(0)) for four channels are set to (0.45, 0.98), (0.55,

0.98), (0.55, 0.98), and (0.55, 0.98), respectively, according to thevolume overlaps of the segmentation maps between the SPGRimage and the DTI image (Liu et al., 2006). The prior probabilityof f(Yi) is available in the single-channel segmentation usingHMRF-EM (Zhang et al., 2001). In this experiment, we averagethe probability maps of the four channel segmentations with equalweight to obtain f(Yi).

As an example, Table 1 provides the results for a randomlyselected case. It is clear from the results that the CSF and non-CSFsegmentation results from the four channels are quite different,e.g., the CSF volume percentages vary from 11.7% to 22.2% whilethe CSF percentage on SPGR data is 19.7%. The CSF percentageobtained by multi-channel fusion is 21.4%, which is very close tothat of the SPGR segmentation result. Considering that tissuesegmentation based on SPGR image is relatively accurate, theseresults indicate the relatively good performance of the proposedtissue segmentation method.

For the fusion of WM and non-WM segmentations, the WMpercentages obtained by different channels are also quite different,

WM (%) Oj

(%)Aj

(%)pj, qj (%)

39.3

%)

45.2 74.4 50.4, 99.956.3 96.0 67.1, 99.654.3 88.0 65.6, 99.960.7 94.0 96.1, 99.970.5 95.8

90.1 73.5 100, 93.487.4 77.4 100, 93.985.2 80.3 100, 94.386.3 88.5

he third and fourth columns show the volume percentages of each tissue types

j), which is determined as the tissue volume overlap ratio of each DWI/DTIws volume agreement (Aj, please refer to Liu et al., 2006 for the definition ofel of every single channel.

118 T. Liu et al. / NeuroImage 38 (2007) 114–123

ranging from 58.6% to 67.6%. Similarly, the multi-channel fusionrenders closer result (49.5%) to the SPGR segmentation result(39.3%). It should be noted that there are relatively large gapsbetween the fusion result and the SPGR segmentation result.However, using the multiple-channel fusion method, these gapscan be improved.

The results in this section support our point in the Multi-channelfusion subsection that the single-channel segmentation is less re-liable, and that an optimal combination of multiple channels, e.g.,using the STAPLE algorithm, provides improved performance.

Experiment 2

This experiment provides an example of evaluation by visualinspection. As shown in Fig. 3, it is difficult to obtain an idealtissue segmentation result by using only one channel in DTI space.For example, in the segmentation of CSF/non-CSF (Fig. 3a), theCSF volume agreement (please refer to Liu et al., 2006 for thedefinition of volume agreement) with that of SPGR segmentationfor the ADC channel is 74.4%. Using our multi-channel fusionmethod, the CSF volume agreement increases to 95.8%. Anotherexample is that in the segmentation of WM and non-WM regions(Fig. 3b), the WM volume agreement of the FA channel with thatof SPGR segmentation is 73.5%. On the other hand, our multi-channel fusion method increases the agreement to 88.5%.

Fig. 3. Multi-channel data fusion. (a) CSF tissue maps obtained by different methodtop is the segmentation by simple voting strategy. The right bottom is the result by 7The left four columns are results by single-channel segmentation. The right two b

In general, the multi-channel fusion method based on DTI dataobtains visually reasonable tissue segmentation results, comparedto the tissue segmentation results based on SPGR data (Fig. 3).This supports our claim about the advantage of multi-channelfusion introduced in the Multi-channel fusion subsection. Ourexperimental result shows that it is possible to obtain a reasonablygood tissue segmentation map by combining the multiple-channelsegmentation results based on the DTI data only.

To compare the proposed multiple-channel fusion method withthe simple voting method, Fig. 3(a) shows the voting result ofADC, λ1, λ2, and λ3 as an example. The CSF volume percentageof the segmentation result by voting is 13.5%, which is far awayfrom the SPGR segmentation result (19.7%). However, the multi-channel fusion method obtains more consistent segmentationresult, compared to the SPGR channel segmentation.

Experiment 3

We have evaluated the tissue segmentation method based onDTI data for 10 cases. Fig. 4 shows the volume overlaps andagreements between DTI segmentation and SPGR segmentationfor these 10 cases (please refer to Liu et al., 2006 for the definitionsof volume overlap and volume agreement).

Note that in Fig. 4(a) the volume overlaps for the differentchannels are quite discrepant, with different cases having different

s. The left five columns are results by single-channel segmentation. The right-channel fusion. (b) WM and GM tissue maps obtained by different methods.ottom columns are results obtained by 7-channel fusion.

Fig. 4. Evaluation of the tissue segmentation method. (a) Volume overlap for individual channel segmentation result and that for fusion result. The averageoverlaps of four single-channel segmentations and the fusion result are 0.44±0.08, 0.48±0.13, 0.60±0.08, 0.57±0.07, and 0.63±0.04, respectively. (b) Volumeagreement for individual channel segmentation result and that for fusion result. The average agreements of four single-channel segmentation results and thefusion result are 0.66±0.18, 0.84±0.23, 0.88±0.05, 0.92±0.08, and 0.94±0.04, respectively.

119T. Liu et al. / NeuroImage 38 (2007) 114–123

channels that yield the highest level of overlap. For example, in thefirst case, the λ3 channel has the highest overlap, whereas in thesecond case, the λ2 channel has the highest overlap. This clearlyelucidates the value of the multiple-channel fusion method inestimating the performance levels of different channels and inobtaining the final segmentation result based on more reliablechannels. In general, the fusion result of tissue segmentation hasthe highest volume overlap averaged over the 10 cases, as shown inFig. 4(a). The mean value and the stand deviation of overlaps forfour individual channel segmentations and the fusion result are0.44±0.08, 0.48±0.13, 0.60±0.08, 0.57±0.07, and 0.63±0.04,respectively. This partly shows that the fusion strategy generatesmore desirable tissue segmentation results than a single channelbased on DTI data, considering that tissue segmentation based onSPGR image can serve as a comparison target.

The results of volume agreement for these 10 cases are shown inFig. 4(b). We see that the channels with maximum volumeagreement vary with each case (see Fig. 4b). This again emphasizesthe importance of the multiple-channel fusion method in differ-entiating the performance levels of different channels. Fig. 4(b)shows that the multiple-channel fusion result has the highest volumeagreement averaged over the 10 cases. On average, the mean values

and the standard deviations of agreements for four individualchannel segmentation results and the fusion result are 0.66±0.18,0.84±0.23, 0.88±0.05, 0.92±0.08, and 0.94±0.04, respectively.

Experiment 4

The previous experiments show that the multi-channel fusionalgorithm has better results from only one channel. To demonstratehow the different channels affect the results of multi-channel fusionsegmentation, we evaluate the contribution from different channelsby measuring the change of the posterior probability (reference toEq. (6)) between Wi and Wij, where Wi is the posterior probabilityobtained from all channels according to Eq. (6), and Wij is also theposterior probability obtained from all channels except channel j.

We can define the change weight as following:

CWj ¼

Xi

jWij �WijXj

Xi

jWij �Wijð8Þ

where j indicates the channel we use for the multi-channel fusion,and i indicates the voxels of the image volume. According to Eq.

Fig. 5. Change weight of posterior probability of four channels (ADC, λ1, λ2, λ3) for CSF segmentation over 10 cases. The average change weight of posteriorprobability for the four channels are 0.17±0.07, 0.20±0.04, 0.43±0.08, and 0.19±0.10, respectively.

120 T. Liu et al. / NeuroImage 38 (2007) 114–123

(8), the channel with high change weight (CW) indicates highcontribution because the posterior probability varies greatlywithout this channel, whereas the channel with low change weightindicates low contribution because the posterior probability varieslittle with or without this channel.

We have done this experiment over the 10 cases. Fig. 5 showsthe change weight for four channels (ADC, λ1, λ2, λ3). Theaverage change weight of the posterior probability for the fourchannels are 0.17±0.07, 0.20±0.04, 0.43±0.08 and 0.19±0.10,respectively. Based on the result in Fig. 5, it is apparent that thechannel λ2 has larger contribution to the fusion procedure in mostcases. Nevertheless, the other three channels also have importcontributions to the segmentation result.

Experiment 5

In this section, we compare the proposed multiple-channelfusion method with our previous method of using two only, forinstance, ADC and FA channels (Liu et al., 2006), as well as withthe multi-spectral segmentation algorithm in the FSL MFAST(Zhang et al., 2001). It is noted that when performing the multi-spectral segmentation, we used all of the seven channels. Toquantify the comparison, Table 2 shows the volume overlaps (Oj)and volume agreements (Aj) between the segmentation maps using

Table 2Volume overlap (Oj) and volume agreement (Aj) with different algorithms

WM GM CSF

Oj (%) Aj (%) Oj (%) Aj (%) Oj (%) Aj (%)

ADC+FA 0.83 0.84 0.56 0.94 0.28 0.56Multi-spectral 0.46 0.74 0.43 0.84 0.80 0.57Multi-channel 0.68 0.89 0.64 0.85 0.63 0.94

The result is averaged over 10 cases. The third, fourth, and fifth rows showthe volume overlaps and volume agreements of the segmentation resultsusing the ADC+FA channel, using the FSL MFAST multi-spectral seg-mentation algorithm, and using the proposed multi-channel fusion algo-rithm, respectively.

different algorithms with that obtained by the SPGR segmentationmap. It is apparent that the results using the ADC and FAchannels have much lower CSF overlaps and agreements than themulti-spectral segmentation algorithm and the proposed multiple-channel fusion method. It is noted that the volume overlaps andagreements of CSF are much lower than the results in Liu et al.(2006) because the SPGR segmentation result was used to guidethe segmentation of CSF in Liu et al. (2006), but not used in thiswork. Because low volume overlap and agreement in CSF wouldcause severe problems of heterogeneous voxels in the measure-

Fig. 6. (a) Tissue segmentation map using ADC and FA channels. (b) Tissuesegmentation map using FSL MFAST multi-spectral segmentation algo-rithm. (c) Tissue segmentation map using multiple-channel fusion algorithm.(d) Tissue segmentation map using SPGR image.

121T. Liu et al. / NeuroImage 38 (2007) 114–123

ment of gray matter properties, the result in Table 2 verifies thatthe proposed multiple-channel fusion method is much morepreferred than the tissue segmentation method based on only twochannels.

For the multi-spectral segmentation algorithm, although its per-formance in CSF volume overlap and agreement is comparable tothe proposed multiple-channel fusion method (Table 2), its perform-ances in GM and WM volume overlaps and agreements are poor(Table 2). The reason might be that the proposed multiple-channelfusion method exploits the STAPLE algorithm to estimate theoptimal combination of the segmentations in seven different chan-nels, whereas the multi-spectral segmentation algorithm does not.

As an example, Fig. 6 shows the segmentation results using thesegmentation method based on the two channels of ADC and FA,the segmentation result using the FSL MFAST multi-spectralsegmentation method, and the segmentation result using the pro-posed multi-channel fusion method and the SPGR segmentationresult. Clearly, the proposed multiple-channel fusion method gene-rates closer segmentation result with SPGR segmentation resultthan the other two methods.

Application: alternative brain tissue segmentation method

Tissue segmentation based on structural data, e.g., T2-weightedimage, sometimes can produce undesirable segmentation result.For example, tissue segmentation using T2-weighted image cannotaccurately distinguish the putamen and the thalamus, as shown inFig. 7(b), whereas tissue segmentation based on DTI data canachieve better results, as shown in Fig. 7(e) (The data are provided

Fig. 7. Tissue segmentation using T2-weighted image and DTI images (seven chanADC image. (d). FA image. (e). Tissue segmentation map through multi-channelspace has better results than that in T2-weighted image.

by Dr. Susumu Mori of the Johns Hopkins University. Detailsabout imaging settings are referred to http://cmrm.med.jhmi.edu/).White arrows point to putamen and thalamus areas where tissuesegmentation in DTI space has better results than that based on T2-weighted image. Therefore, in the absence of structure data or inthe case where desirable segmentation result from structure datacannot be obtained, tissue segmentation based on DTI data via themultiple-channel fusion method provides an alternative means toobtain tissue maps of the brain.

Discussion and conclusion

We have demonstrated that the brain tissue segmentation basedonly on single channel in DTI space produces less reliable result, andthat the multiple-channel fusion method presented in this paper cansubstantially improve the segmentation. There are two importantaspects to the final segmentation results: (1) the performance ofindividual channel segmentation and (2) the assessment of indi-vidual channel segmentation and fusion of them. A variety of issuesare related to the first aspect, including the performance of individualchannel segmentation, e.g., DTI data quality and the segmentationmethod used. In the future, we will test our multiple-channel fusionmethod on datasets with various image resolutions and qualities. Inaddition, we will test different tissue segmentation methods, e.g., themethods in Wells et al. (1996), Pham and Prince (1999), Dale et al.(1999), for individual channel segmentation, and investigate howdifferent segmentation methods affect the final result.

The second issue is addressed by the published STAPLEalgorithm, which identifies the performance levels of different

nels). (a) T2-weighted image. (b) Tissue segmentation using T2 image. (c).fusion. White arrows point to areas where tissue segmentation in DWI/DTI

122 T. Liu et al. / NeuroImage 38 (2007) 114–123

segmentation maps and estimates an optimal combination of thesegmentation. This is fundamentally different from the voting rule(Warfield et al., 2004). In this paper, we used one fixed set ofoverlaps (please refer to subsection 3.2.) for the initialization of(pj

(0), qj(0)). In our future work, we will investigate how the

initialization of (pj(0), qj

(0)) will influence the final multi-channelfusion results, and how to achieve desirable initializations.

In this work, we use seven channels that are the mostfrequently used in the DWI/DTI image analysis, as we believethese channels provide complementary important information,e.g., the λ1, λ2, and λ3 channels describe the diffusion of threeorthogonal directions separately. However, how many channels, orwhat combination of these channels, is adequate to obtain asatisfactory brain tissue segmentation result would need furtherinvestigation. Another interesting issue to be investigated is howthe inclusion of channels that are not intrinsically within the samespace as DTI, e.g., T1- or T2-weighted channel, into the STAPLEfusion procedure will influence the final multi-channel fusionresult.

The STAPLE algorithm assumes that different segmentationsare conditionally independent, and then estimates an optimalcombination of these segmentations. In this paper, some channelsare derived from the three independent eigenvalues and are notmutually independent. How this mutual dependence betweendifferent channels influences the STAPLE fusion result remainsunclear. This issue will be investigated in our future work. It is alsonoted that, in our application, all of the seven channels areindependently segmented into two class maps, which does addrandomness and independence into the final segmentation mapsthat are used as inputs to the STAPLE algorithm. But what degreeof independence this separate two-class segmentation has addedremains unclear. Future work to investigate the degree ofdependence between the seven channels, as well as between thesegmented maps, would be valuable.

In the absence of digital DTI phantoms, currently, we evaluatethe proposed tissue segmentation method by measuring the volumeagreement and volume overlap between the segmentation results inDWI/DTI space and that in SPGR space. A similar assumptionmade here as that in Liu et al. (2006) is that the SPGR imageprovides relatively reliable tissue contrast and segmentation.However, it is noted that SPGR images do not have ideal tissuecontrast and cannot serve as a gold standard. Voxel-based com-parison using the volume overlap measure might be inaccurategiven the possible misalignment between the DTI image and SPGRimage caused by a variety of reasons such as the geometricdistortion in DTI imaging, the partial volume effect, the reslicingand interpolation of DTI data, and the co-registration error. In thefuture, evaluation and validation studies could be performed basedon comparison of the automated segmentation result with thoseobtained by expert manual segmentation results.

In summary, we presented a brain tissue segmentation methodbased on DTI data. This method fuses seven two-class segmentationmaps, which are generated by utilizing the tissue contrast exiting inthe corresponding single channel, to obtain a complete brain tissuesegmentation map using the STAPLE algorithm. The tissuesegmentation results for 10 test cases show that the single-channelsegmentation in DTI space is less reliable, and an optimal com-bination of seven selected channels produces significantly improvedresults. The STAPLE algorithm also plays a key role in computingthe probabilistic estimate of the true segmentation and a measure ofthe performance level represented by each segmentation.

Acknowledgments

This research was funded by a research grant to STCW byHarvard Center for Neurodegeneration and Repair, HarvardMedicalSchool. Parts of public DTI and SPGR datasets from NAMIC wereprovided by the Laboratory of Neuroscience, Department ofPsychiatry, Boston VA Healthcare System and Harvard MedicalSchool, which is supported by the following grants: NIMH R01MH50740 (Shenton), NIH K05 MH01110 (Shenton), NIMH R01MH52807 (McCarley), NIMH R01 MH40799 (McCarley), VAMerit Awards (Shenton; McCarley), and VAResearch EnhancementAward Program (REAP: McCarley). We want to express our thanksto Dr. Susumu Mori for sharing the DTIStudio software and DTIdatasets, to STAPLE authors and ITK for sharing the STAPLE filter,and to FSL developers.

References

Bammer, R., 2003. Basic principles of diffusion-weighted imaging. Eur. J.Radiol. 45, 169–184.

Basser, P.J., Jones, D.K., 2002. Diffusion tensor MRI: theory, experimentaldesign and data analysis—A technical review. NMR Biomed. 14,456–467.

Basser, P.J., Mattiello, J., LeBihan, D., 1994. Estimation of the effectiveself-diffusion tensor from the NMR spin echo. J. Magn. Reson. 103,247–254.

Dale, A.M., Fischl, Bruce, Sereno, M.I., 1999. Cortical surface-basedanalysis: I. Segmentation and surface reconstruction. NeuroImage 9 (2),179–194.

Gonza’lez Ballester, M.A., Zisserman, A., Brady, M., 2002. Estimation ofthe partial volume effect in MRI. Med. Image Anal. 6 (4), 389–405.

Helenius, J., Soinne, L., Perkio, J., Salonen, O., Kangasmaki, A., Kaste,M.R., Carano, A.D., Aronen, H.J., Tatlisumak, T., 2002. Diffusionweighted MR imaging in normal human brains in various age groups.AJNR Am. J. Neuroradiol. 23 (2), 194–199.

Jezzard, P., Balaban, R.S., 1995. Correction for geometrical distortion inecho planar images from Bo field variations. Magn. Reson. Med. 34,65–73.

Johanna, H., Lauri, S., et al., 2002. Diffusion-weighted MR imaging innormal human brains in various age groups. AJNR 23, 194–199.

Kubicki, M., Park, H., Westin, C.F., Nestor, P.G., Mulkern, R.V., Maier,S.E., Niznikiewicz, M., Connor, E.E., Levitt, J.J., Frumin, M.,Kikinis, R., Jolesz, F.A., McCarley, R.W., Shenton, M.E., 2005. DTIand MTR abnormalities in schizophrenia: analysis of white matterintegrity. NeuroImage 26 (4), 1109–1118 (Jul 15).

Le Bihan, D., 1991. Molecular diffusion nuclear magnetic resonanceimaging. Magn. Reson. Q. 7, 1–30.

Le Bihan, D., Mangin, J.F., Poupon, C., et al., 2001. Diffusion tensorimaging: concepts and applications.Magn. Reson. Imaging 13, 534–546.

Li, S.Z., 2001. Markov Random Field Modeling in Image Analysis.Springer-Verlag, New York.

Liu, T., Young, G.S., Chen, N.K., Huang, L., Wong, S.T.C., 2005. 76-Spaceanalysis of gray matter diffusivity: methods and application. MICCAI2005, Palm Springs, California.

Liu, T., Young, G., Huang, L., Chen, N.K., Wong, S.T.C., 2006. 76-Spaceanalysis of grey matter diffusivity: methods and applications. Neuro-Image 31 (1), 51–65 (May 15).

Moseley, M.E., Kucharczyk, J., Mintorovitch, J., et al., 1990. Diffusion-weighted MR imaging of acute stroke: correlation with T2-weighted andmagnetic susceptibility-enhanced MR imaging in cats. AJNR 11,423–429.

Perona, P., Malik, J., 1990. Scale-space and edge detection using anisotropicdiffusion. IEEE Trans. Pattern Anal. Mach. Intell. 12, 629–639.

Pham, D.L., Prince, J.L., 1999. Adaptive fuzzy segmentation of magneticresonance images. IEEE Trans. Med. Imag. 18 (9), 737–752.

123T. Liu et al. / NeuroImage 38 (2007) 114–123

Sundgren, P.C., Dong1, Q., Gómez-Hassan1, D., Mukherji1, S.K., Maly1,P., Welsh1, R., 2004. Diffusion tensor imaging of the brain: review ofclinical applications. Neuroradiology 46 (5), 339–350.

Warfield, S.K., Zou, K.H., et al., 2004. Simultaneous truth and performancelevel estimation (STAPLE): an algorithm for the validation of imagesegmentation. IEEE Trans. Med. Imag. 23, 903–921.

Wells, W.M., Grimson, E.L., Kikinis, R., Jolesz, F.A., 1996. Adaptivesegmentation of MRI data. IEEE Trans. Med. Imag. 15, 429–442.

Woods, R.P., Grafton, S.T., Watson, J.D.G., Sicotte, N.L., Mazziotta,J.C., 1998. Automated image registration: II. Intersubject validationof linear and nonlinear models. J. Comput. Assist. Tomogr. 22,153–165.

Zhang, Y., Brady, M., Smith, S., 2001. Segmentation of brain MRimages through a hidden Markov random field model and theexpectation-maximization algorithm. IEEE Trans. Med. Imag. 20 (1),45–57.