You are on page 1of 11

176

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 16, NO. 2, APRIL 1997

Statistical Approach to Segmentation of Single-Channel Cerebral MR Images


Jagath C. Rajapakse,* Member, IEEE, Jay N. Giedd, and Judith L. Rapoport
Abstract A statistical model is presented that represents the distributions of major tissue classes in single-channel magnetic resonance (MR) cerebral images. Using the model, cerebral images are segmented into gray matter, white matter, and cerebrospinal uid (CSF). The model accounts for random noise, magnetic eld inhomogeneities, and biological variations of the tissues. Intensity measurements are modeled by a nite Gaussian mixture. Smoothness and piecewise contiguous nature of the tissue regions are modeled by a three-dimensional (3-D) Markov random eld (MRF). A segmentation algorithm, based on the statistical model, approximately nds the maximum a posteriori (MAP) estimation of the segmentation and estimates the model parameters from the image data. The proposed scheme for segmentation is based on the iterative conditional modes (ICM) algorithm in which measurement model parameters are estimated using local information at each site, and the prior model parameters are estimated using the segmentation after each cycle of iterations. Application of the algorithm to a sample of clinical MR brain scans, comparisons of the algorithm with other statistical methods, and a validation study with a phantom are presented. The algorithm constitutes a signicant step toward a complete data driven unsupervised approach to segmentation of MR images in the presence of the random noise and intensity inhomogeneities. Index Terms Cerebrospinal uid, nite Gaussian mixture, gray matter, magnetic resonance imaging, Markov random eld, white matter.

I. INTRODUCTION RAIN matter, as assessed by magnetic resonance (MR) imaging, can generally be categorized as white matter, gray matter, cerebrospinal uid (CSF) or vasculature. Because most brain structures are anatomically dened by boundaries of these tissue classes, a method to segment tissues into these categories is an important step in quantitative morphology of the brain. An accurate segmentation technique may facilitate detection of various pathological conditions affecting brain parenchyma [1][6], radiotherapy treatment and planning, surgical planning and simulations [7], and three-dimensional (3D) visualization of brain matter for diagnosis and abnormality detection [8], [9]. It has already been useful in studying brain development [10][12] and human aging [13], [14].
Manuscript received January 18, 1996; revised September 11, 1996. The Associate Editor responsible for coordinating the review of this paper and recommending its publication was T. Taxt. Asterisk indicates corresponding author. *J. C. Rajapakse is with the Child Psychiatry Branch, National Institute of Mental Health, Bldg. 10, Rm 6N240, 10 Center Drive, MSC 1600, Bethesda, MD 20892-1600 USA (e-mail: jcr@helix.nih.gov). J. N. Giedd and J. L. Rapoport are with the Child Psychiatry Branch, National Institute of Mental Health, Bethesda, MD 20892-1600 USA. Publisher Item Identier S 0278-0062(97)02406-3.

Automated segmentation and morphological measurements of brain tissues and structures may be a prelude to automatic identication of brain structures and functional divisions of the cortex, and complementary to brain functional studies [15], [16]. Even with the recent progress of computer image analysis techniques and MR technology, and considerable interest in clinical studies, an automated method for segmentation of brain tissue is still not available for clinical use. The process of MR imaging produces intensities depending and relaxation on three tissue characteristics, namely: times and proton density (PD). The effect of these parameters on the images can be varied by adjusting parameters such as time to echo (TE) and time to repeat (TR) of the pulse sequence. By using different parameters or numbers of echoes in the pulse sequence, a multitude of nearly registered images with different characteristics of the same object can be achieved. If only a single MR image of the object is available, such an image is referred to as single-channel (single-echo) and when a number of MR images of the same object at the same section are obtained, they are referred as multichannel (multispectral or multiecho) images. For a given scanning time, the voxel sizes achieved in the multichannel images are larger than those in single-channel images. Multichannel images provide more information at a voxel site than single-channel images. Because single-channel images yield small voxel sizes, they are suitable for precise and accurate quantitative measurements of tissues and different cerebral structures. Most segmentation techniques have relied on multispectral characteristics of the MR images [17][43] while a few studies have reported segmentation from single-channel images [44][51]. In this report, we present an automated segmentation technique for single-channel cerebral images obtained using an SPGR (spoiled gradient recalled echo) sequence, an earlier version of which has been successfully applied to a database of more than 500 MR brain scans [11]. Currently available methods for MR image segmentation can be categorized into classical, statistical, fuzzy, and neural network techniques [52][54]. Classical techniques include the use of standard image processing techniques such as thresholding [17][19], [44], and edge- and region-based techniques [45], [46]. Since classical methods do not employ a priori information, the nal segmentations are sensitive to noise and usually do not result in continuous regions. Although neural network techniques have been proposed for MR image segmentation, they have been applied to only a small number of images [20], [21], [47]. Fuzzy segmentation techniques

02780062/97$10.00 1997 IEEE

RAJAPAKSE et al.: SEGMENTATION OF SINGLE-CHANNEL CEREBRAL MR IMAGES

177

have not been applied for single-channel image segmentation although they have shown promise in segmenting multichannel images [22], [23]. We take a statistical approach as has been done by many others for MR image segmentation [24][42], [48][51]. The main obstacles to segmentation of MR images are the presence of noise, errors in the scanners, and the structural variations of the imaging objects, which can be categorized into four different types: thermal/electronic noise, magnetic eld inhomogeneities, biological tissue variations, and partial volume effects. The failure of classical segmentation techniques is due to their inability to cope with these factors. Thermal or electronic noise is random, mostly Gaussian, white, and additive. Since the constituents and their congurations in different tissue types are different, this noise is tissue dependent. Filtering techniques have been used in removing random noise without blurring edges and structural information over the images [55]. Intensity inhomogeneities may result from irregularities in the static and radio frequency (RF) magnetic elds and in the sensitivities of RF receiver coils, and from magnetization of the object being imaged [56]. Magnetic eld inhomogeneities modulate image intensities by slowly varying intensity gradients. Although low-pass and homomorphic ltering techniques have been used to correct the magnetic eld inhomogeneities [17], [28], such procedures may corrupt the edges and other high-frequency details of the image. Measurement of magnetic eld proles using phantoms has been utilized to correct the inhomogeneities [57], [58], but when the images are obtained over a long period of time, a single-phantom prole cannot reliably be used to correct the images. This requires acquisition of a reference-phantom image for every subject scanned. Mathematical modeling techniques assuming that the intensity values for single types of tissues do not vary signicantly over a particular slice [59], [60], and taking more than one scan for the subject [61] are also available for correcting the effect of inhomogeneities. In addition to magnetic eld inhomogeneities, biological variations in different structures belonging to the same tissue class may affect the tissue classication. In our approach, we incorporate the effect of inhomogeneities and biological variations of tissues as variations of the parameters of the statistical model and estimate these parameters at each site using local information. Although the present segmentation algorithm uses signal intensities at the voxel sites as input features to the segmentation process, it can easily be extended when other features such as textural and higher-order statistical features of the intensities are available [50], [62], [63]. When multispectral images are used, the intensity vectors or MR tissue parameters and PD can be used as input features. Although MR tissue parameters are reproducible and less sensitive to noise and inhomogeneities [64][66], it is not possible to derive them from single-channel images. An advantage of using image intensity as an input feature is that it carries information rendered by all MR characteristics of the tissue class. Because of the large number of MR scans and of voxels (256 256 124) in a single brain scan in our database, the optimal segmentation technique should be unsupervised

and automated. Also, the segmentation should be done in a reasonable amount of time. With a large number of slices, manual techniques may be tedious and time-consuming. Because supervised or semiautomated methods are not appropriate for our purpose, we will focus only on unsupervised methods of segmentation. Our algorithm is based on a statistical model that consists of a measurement model and a prior model. The measurement model relates the image data to the segmentation and takes the form of a nite Gaussian mixture [67], [68]. The prior model takes into account smoothness and piecewise contiguous nature of the tissue regions and is modeled by a 3-D Markov random eld (MRF) [69][76]. Our segmentation scheme differs from similar statistical approaches for MR image segmentation used by others [24][42], [48][51] in the following ways: 1) an accurate segmentation is obtained using a three-tissue model (gray matter, white matter and CSF), by extracting the cerebrum from the head scan prior to the segmentation; 2) a 3-D volume segmentation is performed instead of slice by slice segmentation; 3) the method accounts for magnetic eld inhomogeneities and biological variations of the tissues; 4) the model parameters are completely estimated using the input image data. Our measurement model assumes that image intensities of tissue classes are corrupted by additive and white Gaussian noise, as assumed by all others taking the statistical approach except one study [49] which assumed skewed Gaussian distributions to count for the partial volume effect. Since each tissue consists of different types of constituents, generally the noise is tissue dependent. Only one recent study [42] incorporated the biological variations of tissues and eld inhomogeneities into the statistical model. In our statistical model, the biological variations of a particular tissue class are accounted for by assuming that the mean intensities of the tissue classes are slowly varying spatial functions. The magnetic eld inhomogeneities modify both mean tissue intensities and noise variances in a similar manner. The algorithm is adaptive in the sense that the means of the intensity distributions and noise variances are estimated at every voxel site using the local intensity information. Partial volume effects are not considered in our model although the present model can be easily extended to include partial volume effects in a manner similar to approaches taken by others [39], [48]. Previous statistical approaches [24][36], [48], [49] have considered only the measurement part of the statistical model as their complete model and used maximum likelihood (ML) estimation [77] to obtain the nal segmentation. Most used the expectation-maximization (EM) algorithm [67], [68], [78] to compute the ML estimates of the parameters. The outcome of such a method can be very sensitive to noise, and the piecewise contiguous nature of the tissue regions may be easily violated. To overcome this, recent studies have attempted to use MRF to model smooth piecewise continuous regions of the tissue classes [37][41], [51], [79]. The MRF prior model is suited because it connes interactions to local neighborhoods and, hence, is computationally tractable. The MRF parameters were estimated in an ad hoc manner in previous studies, and in this

178

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 16, NO. 2, APRIL 1997

study all MRF parameters were estimated using image data. With the measurement and prior model completely known, the algorithm is designed to nd an approximation to the maximum a posteriori (MAP) estimation [72], [80][82] as the nal segmentation. The proposed algorithm iteratively nds a suboptimal segmentation while estimating the model parameters. It will also be shown how a simulated annealing (SA) technique can be utilized to nd the optimal segmentation. In Section II, we present the statistical model that can be used to segment any number of tissue classes. Section III describes an iterative algorithm which approximately evaluates the MAP estimation of the segmentation and determines the parameters of the model. The results of the segmentation algorithm to a sample of MR brain scans are presented in Section IV. The comparison of our results with other statistical schemes is also presented. In Section V we discuss the advantages and disadvantages of the present method. II. STATISTICAL MODEL We consider a 3-D image volume that consists of a set of contiguous anisotropic rectangular parallelepipeds referred to as voxels. The sites of the voxels are dened on a 3-D cubic grid of points and indexed by Cartesian coordinates. The set of coordinates of the voxel sites in the image is denoted by and the image is denoted by where denotes the image intensity at the voxel site indexed by . In this section, we assume that the spatial volume of each voxel is occupied by only one type of tissue class and leave the effect of partial voluming for the discussion. Segmentation of the image is an assignment of the correct tissue class to every voxel in the image. Let us assume that the total number of tissue classes in the image is and each tissue class is represented by a label from . If denotes an instance of a random variable representing the tissue class at the voxel site , we write to indicate that the tissue class is assigned to the voxel site . A segmentation is then denoted by . The process of segmentation is to nd which represents the correct tissue class at each voxel site given by image . We attempt to nd the MAP estimation from the image data [81]. That is, if represents the optimal segmentation (1) is where is the set of all possible segmentations and the posterior density of the segmentation given the image . Because the prior probability of image is independent of the segmentation , from Bayes theorem [77] (2) where is the conditional probability density of the image given the segmentation and is the prior density of the segmentation . Our attempt is to nd the MAP estimate by modeling (the measurement model), and (the prior model). Each tissue class has a signature, or mean intensity, and the image data at a particular site represents a noise-corrupted version of the signature of the tissue class at that voxel.

We assume that the noise is additive, white, Gaussian, tissue dependent, and space variant. With this assumption, our measurement model, if , can be written as (3) and represent the mean image intensity of the where class at site and the noise signal at site for the tissue class , respectively. If the standard deviation of the noise for the tissue class at site is , then the noise distribution is given by (4) The measurement model is characterized by the parameter set , where . The variations of mean and standard deviation of intensities of the tissue classes over the voxel sites account for the biological variations and spatial intensity inhomogeneities. Since the conditional density is modeled by a tissuedependent and space-variant white Gaussian process and no formal model is assumed for tissue parameters, the image data are conditionally independent. If denotes the region or the set of all voxel sites belonging to tissue class , then the conditional density can be written as (5)

(6)

(7) denotes the total number of voxel sites in the image. where The prior model is based on a 3-D MRF and assumes that the intensity prole is piecewise contiguous, and the adjacent voxels are likely to have similar constituents. If denes a neighborhood system over the image domain, the random variable representing is an MRF if for all , and (8) indicates the neighborhood , excluding the site where . On 3-D space, a homogeneous neighborhood system can be introduced considering a ball of xed radius at every voxel site. A neighborhood ball of order at voxel site is dened by (9) is referred to as the radius of the neighborhood. where When the voxels are anisotropic, a cubic neighborhood with each dimension related to the resolution of voxels in each direction is appropriate. Since is assumed to be distributed by an MRF, according to the HammersleyClifford theorem [69], the probability

RAJAPAKSE et al.: SEGMENTATION OF SINGLE-CHANNEL CEREBRAL MR IMAGES

179

density of is given by a Gibbs distribution [72], [82], [83] having the form (10)

where the energy function

(15) where is a normalizing constant and the summation is taken over all the cliques over the image. A clique is a set of points that are neighbors of one another. The potentials of the cliques depend on their congurations. Usually, the exponential of (10) is associated with a constant referred to as temperature, and we have incorporated the temperature in the clique potentials. Because of the suitability for describing piecewise continuous regions in an image, a class of Gibbs distributions known as multilevel logistic (MLL) models [84] is used here to represent the prior model. In the MLL models a parameter is associated with each clique type , irrespective of the class of its constituents, except at one-site cliques. When only interactions between two sites are considered, the potential function for the clique type is given by (11) is a real positive constant and ( ) takes the value where one when the condition inside is satised; zero otherwise. For one-site cliques the potentials are dened as follows: for (12) In short, the optimum segmentation and parameters for the models are needed to evaluate by the segmentation algorithm. The optimal parameters should predict image data with a maximum probability. Given all the parameters for the model, a Gibbs sampler and the SA scheme [72], [83], [84] can be used to nd the minimum energy conguration for . It has been proven in [72] that with known parameters and annealing schedule, the SA procedure converges to a global maximum of the posterior distribution in an innite amount of time. Because of the size of our image data, and the complex computation involved in this process, such a scheme is practically prohibitive. Therefore, a practical and recursive approximate algorithm which converges to a local minimum of while estimating the elements of and is presented below. The algorithm maximizes the conditional density of the image data at each site given the image and the current estimate of at all other sites. This is the iterated conditional modes (ICM) procedure proposed by Besag [80], which iteratively converges to a local minimum of the energy function. It corresponds to instantaneous freezing of the SA procedure. With a good starting conguration, the ICM procedure has shown to give a fairly good segmentation [82]. From the Markovian property and whiteness of noise (18) (19) represent the binding between The rst two terms of the image data and the segmentation; the last term indicates the smoothness constraints in the prior model. The problem of nding the MAP estimate of the segmentation is same as the minimization problem of the energy function . III. SEGMENTATION ALGORITHM Given the image and knowing the parameters and , we need to nd that minimizes the energy function given , given all the in (15). Finding a global minimum of congurations , is a practically difcult task. The difculty is further compounded by the fact that the model parameters necessary for the segmentation are not known. Hence, we attempt to nd the optimal segmentation while estimating the optimal parameters for the segmentation. Therefore, the segmentation problem can be expressed in the following twostep process: (16) (17)

where and is a real constant. reects the a priori knowledge of the relative likelihood of region of tissue types . Since the prior model parameters are assumed to be spatially invariant, the MRF assumed here is homogeneous. For simplicity, we assign nonzero potentials for only one- and two-site cliques and consider only up to second-order neighborhoods for spatial interactions. The MRF is also isotropic, since twosite clique potentials are independent of the orientation of the two sites. If and denote the two-site clique potentials of the rst- and second-order neighbors, respectively, the set . of prior model parameters is There are 6 rst-order neighbors and 12 second-order neighbors at a voxel site in the 3-D image except at the boundaries. We assume free boundaries where the voxel sites near the boundaries have fewer neighbors than those at the interior. Combining (1) and (2) and indicating the dependence of the model on its parameters, if the optimal parameters for the model are known, the optimal segmentation can be written as (13) and are the optimal parameters for the where measurement model and the prior model, respectively. By substituting from (7) and (10) in (2) and omitting the constant factors, we can write the posterior probability (14)

180

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 16, NO. 2, APRIL 1997

where the local energy

is given by

(20) At each site, the energy is computed for all possible classes and the class that minimizes the energy is selected as the tissue class at that site. The tissue class at every voxel site is iteratively determined in order to arrive at the nal segmentation. Before determining the tissue class at a particular site, the measurement model parameters for the site are evaluated using the local information of the current segmentation. A neighborhood is considered at each voxel site , and the ML estimates of the parameters are determined taking means and standard deviations of the intensities of the neighborhood voxels belonging to the class . The size of the neighborhood determines the extent that the algorithm compensates for biological variations of tissues and magnetic eld inhomogeneities over the image. Sometimes, depending on the neighborhood size and location, there may not be enough voxel sites belonging to a particular class in the neighborhood. If the number of voxels in the neighborhood belongs to a particular tissue class (say for class ) is less than a threshold value , the tissue class at that site is undened and cannot be assigned to that tissue class. With a high threshold, more robust computation of parameters is obtained. The threshold may be set considering the dimensions of the voxels, and the size of the smallest possible tissue region. The isolated regions of fewer than voxels disappear during segmentation. Since the computations needed to compute at each site considering a neighborhood are enormous, we compute estimations of parameters only on 3-D grid points separated by appropriate distances and use bilinear interpolation to obtain the intermediate values. The distances between the grid points are set so that the neighborhoods share sufcient overlap for the interpolation. Here, we assume that the spatial distributions of the functions are smoothly varying over the image, and hence the interpolated values are good approximations for the values we would obtain by an exact calculation. A similar approximation scheme was used in the adaptive segmentation algorithm described in [82]. We refer to one iteration as the visitation to a voxel site, estimation of the parameters and determination of the tissue class. The iterations proceed in a raster scan manner on each site in a slice beginning from the rst slice to the nal slice. One cycle of iteration corresponds to visiting all the sites in the image once. The ICM procedure converges to a local minimum of energy function after tissue classes are determined over the image for a few cycles. Convergence is decided when the number of changes of the tissue classes at the voxel sites drops below a certain threshold. At the end of each cycle the parameters of the prior model are estimated. The parameters of the prior model are crucial to

the segmentation as they determine the amount of smoothing of the segmentation. and control the strength of the interactions between the voxel and its rst- and second-order neighbors and govern the tradeoff between the inuence of observables and the prior constraints. A coding method [69] and maximum pseudo-likelihood (MPL) method [80] have been proposed to nd the parameters of MRFs. But when the number of tissue classes is more than two, it is difcult to solve the resulting nonlinear ML equations in these methods. Because of that, we take an approach proposed in [83] to estimate the prior model parameters in . Let us consider a ball of voxels of order two with the center voxel having class and the neighborhood patterns having a conguration . According to the MLL assumptions [(11) and (12)], the joint probability [(10)] of such a pattern is given by (21) and are the total number of class voxels in where the rst-order neighborhood sites excluding the center voxel, and in the second-order neighborhood sites excluding the rstorder neighbors, respectively. By substituting with in (21) for a pattern with the same neighborhood conguration and a center voxel having class , we can obtain (22) The right-hand side of (22) can be estimated using the current estimate of the segmentation . and are obtained using a histogram technique. Assuming and are the number of congurations of patterns and existing over the segmentation, respectively, we can write (23) and , It should be noted that any combination of when and , can be used to obtain a linear equation of parameters in using (22). Since the number of resulting equations is greater than the number of variables, we obtain a least square solution to estimate the prior model parameters using all the congurations for which . To estimate the parameters one needs a realization of the segmentation, and to obtain the segmentation one needs the values of the model parameters. Hence, for the segmentation algorithm to begin, a starting conguration of is needed. Considering only the measurement portion of the model and assuming the parameters of are space invariant, a rough segmentation is initially obtained. With these starting assumptions, the likelihood of the image data can be written as (24) are We further assume that the conditional densities Gaussian. represents the mixing weights of the class in the mixture. The ML estimates of the parameters of the above initial model [(24)] can be computed using the EM algorithm

RAJAPAKSE et al.: SEGMENTATION OF SINGLE-CHANNEL CEREBRAL MR IMAGES

181

[67], [68], [78]. Since the computations involved in the EM algorithm are massive, and it is not necessary to exactly calculate these initial parameters in , the initial segmentation is obtained using the less computationally intensive conditional EM (CEM) algorithm [86]. To give a starting conguration for CEM, the -means algorithm [87] was used with the initial parameters estimated using the mean and standard deviation of the image data, as explained in [11]. The mean values of white matter, gray matter, and CSF are , and , respectively, initialized as and the standard deviations of all tissue classes are initialized to . We found that the real constants worked well for our images and the nal segmentations were insensitive to slight variations of and . The pseudo code for complete segmentation algorithm is as follows: begin Find and for the image data Initialize and spatially invariantly using and Find initial segmentation using the -means and CEM algorithms For initial segmentation, nd s, , and ips LARGE Until (ips threshold) ips 0 For all voxel sites Find and using neighborhoods Find If ( is not the class at ) ips ips 1 Change the class at to Repeat for all Update s, and using segmentation Continue end

TABLE I AVERAGE TISSUE VOLUMES AND RATIOS OF VOLUMES OBTAINED FROM SEGMENTATION OF TEN BRAIN SCANS OF HEALTHY CHILDREN AND ADOLESCENTS FROM AGE 8 TO 14 YEARS. ALL VOLUMES ARE IN CUBIC CENTIMETERS

IV. RESULTS The segmentation algorithm was used to segment a sample of ten brain scans in our database. These healthy subjects ages ranged from 8 to 14 years. All ten scans were obtained on the same GE 1.5 Tesla Signa scanner and interpreted as clinically normal by neuroradiologists in the Clinical Center at the National Institutes of Health, Bethesda, MD. Images with slice thickness of 1.5 mm in the axial plane were obtained using 3-D SPGR sequence in the steady state. SPGR sequences are mainly sensitive to relaxation response times of the imaging tissues. Imaging parameters were: TE 5 ms, TR 24 ms, ip angle 45 , acquisition matrix = 192 256, number of repetitions 1, and eld of view 24 cm. Original 16-bit images were converted to 8-bit images to speed up processing and reduce memory requirements. In all our images the intensities of gray matter, white matter, and CSF in 16-b images fall in the range of values from zero to 255. Therefore, the 8-bit images were conveniently obtained by truncating the 16-bit images without losing any information of the interested tissue classes in the original image.

The brain scans were shelled using an active surface template of the brain to extract the cerebrum. Descriptions about the brain shelling process have been presented in [11] and [88]. Since all scans were interpreted as normal by neuroradiologists who found no evidence of tumors or tissue abnormalities, the remaining matter in the cerebrum was classied as gray matter, white matter, and CSF. Although blood vessels may be classied either as gray matter or white matter, the percentage error resulted from the presence of vasculature is negligibly small. Therefore, the three-tissue model can be accurately used to represent the intensity distributions in our cerebral images. That is, for segmentation, and if and denote gray matter, white matter, and CSF, respectively, we can write . The segmentation algorithm was implemented in C language and executed on a Sparc 10 workstation. The initial segmentation for the algorithm was obtained using the -means algorithm followed by the CEM algorithm with spatially invariant parameters. Since the voxel size for our images was 0.938 0.938 1.5 mm , a cubic neighborhood of 17 17 13 size with was used to compute the measurement model parameters. The convergence was achieved when the number of sites that changed its tissue class in a cycle dropped below 0.01% of the total voxels in the cerebral image. With this threshold, the execution time for one brain was approximately 2 h and the algorithm converged in about 20 cycles of iterations. The nal segmentation was almost unaffected by this threshold since it represented a very small portion of the image volume. Average amounts of gray matter, white matter, and CSF volumes, and gray to white matter volume and CSF to cerebral volume ratios over the ten subjects were computed and are shown in Table I with their standard deviations. The CSF volumes included both ventricular CSF and the extracerebral CSF which were left in the shelling process. The brain stem was separated from the cerebrum at the level of the oor of the third ventricle during the shelling process. The average prior model parameters obtained were , , . It is not possible to nd the absolute values of and s, but the differences are sufcient to determine the tissue classes at each site. A sample of axial slices from a representative brain scan and the corresponding segmented slices are shown in Fig. 1. The slices were taken at lower cortex (temporal lobe section), mid-ventricular section, and upper cortex (lateral ventricular

182

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 16, NO. 2, APRIL 1997

Fig. 1. (a) and (b) Original and (c) and (d) segmented (below) images of a representative brain: (a) lower cortex slice (temporal lobe section), (b) mid ventricular slice, and (c) upper cortex slice (lateral ventricular slice). Fig. 3. Comparison of segmentation with other statistical schemes: (a) using only spatially invariant measurement model and ML estimation as segmentation, (b) using spatially invariant measurement model, a MRF as prior model, and the ICM algorithm for segmentation, (c) segmentation from the presented algorithm, and (d) using the present model but using 250 iterations of simulated annealing (SA) to estimate nal segmentation.

Fig. 2. Illustration of 3-D nature of the segmentation: (a) three-dimensional rendering of an image volume, (b) axial, (c) sagittal, and (d) coronal slices from the segmented brain. Cross lines show the level where the slice were taken.

section). Illustration of the 3-D nature of the segmentation is shown in sagittal, coronal, and axial sections in Fig. 2. A clear continuation of tissues in all planes shows the accuracy of the volume segmentation as opposed to slice by slice segmentation. To compare our results with the other statistical methods, a selected brain was segmented using three other statistical schemes: 1) a nite Gaussian mixture as a spatially invariant model (without prior model) with the EM algorithm used to derive the ML estimations of parameters and segmentation; 2) a spatially invariant Gaussian mixture used as the measurement

model and an MRF as the prior model with parameters of the measurement model obtained using the EM algorithm and the segmentation using the ICM algorithm [80]; 3) using the presented model, the segmentation was obtained by the SA procedure with interruptions to estimate the model parameters in a similar manner described in Section III. The segmentations of the mid-ventricular axial slice using the above schemes and the proposed algorithm are shown in Fig. 3. The space invariant schemes were fast and do not account for intensity inhomogeneities and biological tissue variations. The tested algorithm with the SA procedure [scheme 3] took about two days to converge for a single brain scan. To validate our segmentation a simple phantom using various concentrations of dextran-coated superparamagnetic iron oxide particles (AMI-25 particles, Advanced Magnetics) in agarose gel (Sigma Chemical Co. Type 1, Low EEO) was made based on previous reports [90]. Addition of iron oxide particles decreased relaxation time of the gel resulting in an increase of image intensity in the partially saturated imaging parameters used for this study. After imaging a few test phantoms with 1% (wt/vol) agarose and comparing the intensities with those of the actual tissues, iron oxide concentrations for gray and white matter and CSF were determined. Concentrations of 67 Molar, 33 Molar and 6.7 Molar of iron oxide were selected to represent white matter, gray matter, and CSF, respectively. The phantom was constructed in a cylindrical plastic bottle and imaged parallel to the long axis in the same manner as our subjects. One slice of the phantom image with the segmentation is shown in Fig. 4. Each layer of tissues consisted of 200 ml of agarose with the correct concentration of iron oxide.

RAJAPAKSE et al.: SEGMENTATION OF SINGLE-CHANNEL CEREBRAL MR IMAGES

183

Fig. 4. Segmentation of the phantom: a representative slice from: (a) the phantom image and (b) the segmented image.

The phantom image was segmented, and the volume of each layer was measured and compared with the actual volumes. The error percentages for gray matter, white matter, and CSF volumes were 1.2%, 2.3%, and 2.8%, respectively. V. DISCUSSION
AND

CONCLUSION

Because the image of the cerebrum was extracted from the head scans prior to segmentation, the three-tissue model could conveniently be used in the segmentation process. In earlier approaches, with the skull and muscles intact, the head scan had to be segmented to about seven tissue types. With that number of tissue classes, the risk of misclassication, and instability of segmentation process increase. Therefore, more accurate segmentation results can be expected from using the present segmentation scheme than those which used a higher number of tissue classes. Nonetheless, the present segmentation algorithm was designed to segment any number of tissue classes. The Gaussian assumption for random noise has been used by almost every segmentation study as the measurement model and there is no valid reason to assume any other distribution. Segmentation results obtained from the ML estimation using only the measurement model are noisy and unreal. It is clear that our model is superior to the statistical models that utilize only the measurement part. A prior model is needed to impose the continuity and homogeneity constraints of the tissue regions. Use of the MLL-MRF model gave a reasonable and mathematically tractable prior model. In this report, the MRF was used to model only the intensity process, but incorporation of the discontinuity process into the prior model [72], [81] may improve segmentation results. The energy functions for the discontinuity process in the 3-D images have not yet been studied and further research in this direction is necessary before incorporating 3-D discontinuity processes into the proposed model. Our model incorporates the effect of magnetic eld inhomogeneities and biological variations of tissues as variations of model parameters. Previously proposed statistical methods have not fully incorporated these effects into the statistical model. Such methods have to depend on other techniques to account for these factors. Further, it is difcult to design classical, neural network, or fuzzy algorithms to overcome these effects. With the present statistical approach, it is not necessary to use preprocessing or post-processing of images to remove the effect of biological tissue variations and magnetic eld inhomogeneities. Using tissue-specic and spatially

dependent means and standard deviations as parameters in the measurement model, biological variations and spatial intensity variations are accounted for at the expense of increased computational cost. The tissues of anatomic structures are continuous volumes in 3-D space and therefore, volume segmentation was performed instead of slice by slice segmentation. Slice by slice segmentation was preferred by some researchers because of its ability to overcome inter-slice variations of intensities. But the intensity inhomogeneities exist and should be accounted for in all directions. Since the proposed algorithm considers cubic or spherical neighborhoods when computing local measurement model parameters, it accounts for intensity variations in all directions. Except for the size of the local neighborhood for the computation of the parameters of the measurement model, all other parameters of the model were derived from the image data. The size of the neighborhood was determined by testing the algorithm with different values over the images. Further investigation is necessary to determine this neighborhood size from image data. The computation requirements and the processing time are almost independent of the size of because of the interpolation scheme used in our algorithm. The measurement model parameters were actually computed using neighborhoods for only a limited number of sites and the values for the other sites were interpolated to reduce the massive computational requirement assuming that the intensity variations over the images were slowly varying functions. This is a fairly good assumption since sudden discontinuities of intensities are not expected from good MR imaging scanners. Volume segmentation ensures continuity of the 3-D boundaries of the segmented images whereas slice by slice segmentation does not guarantee continuation of the boundaries of the tissue regions between slices. This is clearly seen in the Fig. 2. When performing slice by slice segmentation, it is necessary to determine the number for tissue classes in each slice since the number of tissue classes varies from one slice to the other, while our volume segmentation approach avoids the necessity of estimating the number of tissue classes in each slice. Special precautions were not taken to overcome partial volume effects present at the tissue boundaries of the images. One major complication in the partial voluming effect is that the combined pixels of white matter and CSF may result in intensities close to gray matter intensities. Because the proposed algorithm does not allow formation of small regions and always favors continuous regions over the jump discontinuities, the chances of formation of thin gray matter boundaries between white matter and CSF region are rare. That is, the prior constraints in the present model partially account for the partial volume effects. The segmented images using our algorithm did not show the presence of partial voluming effect at the tissue boundaries. Still the partial volume effects can introduce statistical errors into the model and the parameter estimation process. Also, since the voxel size of the images used in this study is smaller than those used in multichannel studies, the partial voluming effects resulting in our images are less.

184

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 16, NO. 2, APRIL 1997

In the computations of segmentation, the prior model was restricted to rst- and second-order neighborhoods. By considering higher-order neighborhoods and including all the clique types associated with the neighborhoods, one would expect to get better segmentation provided that the parameter estimation can be accurately and efciently done and the basic model assumption is valid. Because of computation and time limitations, we limited our experiments up to secondorder neighborhoods in the prior model. Since the values obtained in our experiments were close to zero, the of consideration of only rst-order neighborhood seems sufcient. The ICM algorithm which is the basis for our segmentation algorithm, is a deterministic scheme to choose the tissue categories at each voxel site. It converges to a local minimum of the energy function in a small number of iterations. Therefore, the initialization of the segmentation is important and if the initial conguration is far from the optimal conguration, the nal segmentation may end in a local minimum. The proposed initialization scheme gave good initial segmentations and visually correct nal segmentations for our images. It is easily possible to extend our algorithm to include a stochastic type optimization algorithm like SA which is not sensitive to the initialization. But the time and computation requirements to execute such an algorithm are huge. Such schemes are not practical unless access to a parallel computer is available. With our experience with the one brain segmented with SA scheme, stochastic approaches produced slightly better segmentation visually than that obtained with the deterministic schemes (ICM) used in the present study. Although we claim that the segmentation converges to a local minimum of the posterior distribution and estimates of the model parameters are close to optimum, the convergence of the algorithm is not completely known. A proof of the convergence of the segmentation scheme would be reassuring. For all the simulations performed, the algorithm converged in a nite and convenient number of iterations. Although the present algorithm took about 2 h to segment one of our images (124 slices of 256 256 images), the algorithm can be run unattended in batch mode over a large database due to the unsupervised nature of the algorithm. The shelling process with active surface templates sometimes left patches of dura and eye balls which were edited by experienced raters. It sometimes shaved the cortical gray matter and removed the CSF in sulci, which could not be recovered. These artifacts affect the nal volumes of gray matter and CSF volumes. The high variation seen in the CSF volumes may be due to these effects and poorly dened subrachnoid CSF spaces in the shelling process. Our segmentation scheme segmented cortical tissues well, but poorly segmented some of the small subcortical nuclei such as putamen and caudate. The gray level variations due to noise and biological phenomena and the smoothing properties of the MRF model seems to affect the segmentation of subcortical nuclei. One solution to this may be the segmentation of the cortical and subcortical regions separately. The phantom study showed that the errors of the segmentation were less than 3% of tissue volumes. However,

because of the simple shape and homogeneous nature of the representative tissues in the phantom, the partial volume effects and biological tissue variations seen in the brain were not replicated in the phantom. A complex phantom is required to mimic the actual tissue characteristics of the brain. Small voxel sizes in single-channel images facilitate accurate segmentation and quantication of small structures of the brain, and result in less partial voluming errors compared to multichannel images. However, multichannel images play a vital role in brain imaging because they carry information from images responding to different tissue characteristics. Multichannel images are used to design useful segmentation schemes such as to overcome partial volumes effects [91] and to detect abnormalities in tissues. The present algorithm can be extended to segment multichannel images using multidimensional densities for the measurement model. This report presented a statistical model which incorporated major errors that are encountered in practical MR scanners. Use of the three-tissue model increased the accuracy and stability of the segmentation process. The segmentation results are sufciently accurate for using the present algorithm for many clinical applications in the presence of noise and intensity inhomogeneities of the scanners. An earlier fast version of this algorithm has been successfully applied to more than 500 brains of children and adolescents [11] and documented differential change in gray and white matter volumes between ages 4 to 18. The present algorithm will be used to evaluate regional measurements of gray and white matter and agerelated changes as well as comparison of tissue volumes for normals and patients with childhood onset neuropsychiatric disease. REFERENCES
[1] C. R. Jack, Brain and cerebrospinal uid volume: Measurement with MR imaging, Radiol., vol. 178, pp. 2224, 1991. [2] R. B. Zipursky, K. O. Lim, E. V. Sullivan, B. W. Brown, and A. Pfefferbaum, Widespread cerebral gray matter volume decits in schizophrenia, Arch. Gen. Psych., vol. 49, pp. 195205, 1992. [3] T. E. Schlaepfer, G. J. Harris, A. Y. Tien, L. W. Peng, S. Lee, E. B. Federman, G. A. Chase, P. E. Barta, and G. D. Pearlson, Decreased regional cortical gray matter volume in schizophrenia, Amer. J. Psych., vol. 151, no. 6, pp. 842848, 1994. [4] A. Pfefferbaum, K. O. Lim, R. B. Zipursky, D. H. Mathalon, M. J. Rosenbloom, B. Lane, C. N. Ha, and E. V. Sullivan, Brain gray and white matter volume loss accelerates with aging in chronic alcoholics: A quantitative MRI study, Alcoholism Clin. Exp. Res., vol. 16, no. 6, pp. 10781089, 1992. [5] W. L. Olsen, F. M. Longo, C. M. Mills, and D. Norman, White matter disease in AIDS: Finding at MR imaging, Neuroradiol., vol. 169, pp. 445448, 1988. [6] H. Rusinek, M. J. de Leon, A. E. George, L. A. Stylopoulos, R. Chandra, G. Smith, T. Rand, M. Mourino, and H. Kowalski, Alzheimer disease: Measuring loss of cerebral gray matter with MR imaging, Neuroradiol., vol. 178, pp. 109114, 1991. [7] K. Fitzgerald, Medical electronics, IEEE Spectrum, vol. 28, No. 1, pp. 7678, 1991. [8] R. A. Robb, Three-Dimensional Biomedical Imaging. New York: VCH, 1995. [9] M. E. Brummer, R. M. Mersereau, R. L. Eisner, and R. R. J. Lewine, Automatic detection of brain contours in MRI data sets, IEEE Trans. Med. Imag., vol. 12, no. 2, pp. 153166, 1993. [10] A. K. H. Miller, R. L. Alston, and J. A. N. Corsellis, Variation with age in the volumes of gray and white matter in the cerebral hemispheres of man: Measurement with an image analyzer, Neuropathol., Appl. Neurobiol., vol. 6, pp. 119132, 1980.

RAJAPAKSE et al.: SEGMENTATION OF SINGLE-CHANNEL CEREBRAL MR IMAGES

185

[11] J. C. Rajapakse, J. N. Giedd, C. DeCarli, J. W. Snell, A. McLaughlin, Y. C. Vauss, A. L. Krain, S. D. Hamburger, and J. L. Rapoport, A technique for single-channel MR brain tissue segmentation: Application to a pediatric study, Magn. Reson. Imag., vol. 18, no. 4, 1996. [12] T. Autti, R. Raininko, S. L. Vanhanen, M. Kallio, and P. Santavuori, MRI of the normal brain from early childhood to middle age, Neuroradiol., vol. 36, pp. 644648, 1994. [13] K. O. Lim, R. B. Zipursky, M. C. Watts, and A. Pfefferbaum, Decreased gray matter in normal aging: An in vivo magnetic resonance study, J. Gerontol.: Biolog. Sci., vol. 47, no. 1, pp. B26-30, 1992. [14] D. G. M. Murphy, C. DeCarli, M. B. Schapiro, S. I. Rapoport, and B. Horwitz, Age-related differences in volumes of subcortical nuclei, brain matter, and cerebrospinal uid in healthy men as measured with magnetic resonance imaging, Arch. Neurol., vol. 49, pp. 839845, 1992. [15] X. Hu, N. Alperin, D. N. Levin, K. K. Tan, and M. Mengeot, Visualization of MR angiographic data with segmentation and volume-rendering techniques, J. Magn. Reson. Imag., vol. 1, pp. 539546, 1991. [16] D. N. Levin, X. Hu, K. K. Tan, S. Galhotra, C. A. Pelizzari, G. T. Y. Chen, R. N. Beck, C. T. Chen, M. D. Cooper, J. F. Mullan, J. Hekmatpanah, and J. P. Spire, The brain: Integrated three-dimensional display of MR and PET images, Radiol., vol. 172, pp. 783789, 1989. [17] K. O. Lim and A. Pfefferbaum, Segmentation of MR brain images into cerebrospinal uid spaces, white and gray matter, J. Comput. Assist. Tomogr., vol. 13, no. 4, pp. 588593, 1989. [18] G. J. Harris, E. H. Rhew, T. Noga, and G. D. Pearison, User-friendly method for rapid brain and CSF volume calculation using transaxial MRI images, Psych. Res.: Neuroimag., vol. 40, pp. 6168, 1991. [19] G. J. Harris, P. E. Barta, L. W. Peng, S. Lee, P. D. Brettschneider, A. Shah, J. D. Henderer, T. E. Schlaepfer, and G. D. Pearison, MR volume segmentation of gray matter and white matter using manual thresholding: Dependence on image brightness, Amer. J. Neuroradiol., vol. 15, pp. 225230, 1994. [20] S. C. Amartur, D. Piraino, and Y. Takefuji, Optimization neural networks for the segmentation of magnetic resonance images, IEEE Trans. Med. Imag., vol. 11, no. 2, pp. 215220, 1992. [21] M. Ozkan, B. M. Dawant, and R. J. Maciunas, Neural-networkbased segmentation of multi-modal medical images: A comparative and prospective study, IEEE Trans. Med. Imag., vol. 12, no. 3, pp. 534544, 1993. [22] M. E. Brandt, T. P. Bohan, L. A. Kramer, and J. M. Fletcher, Estimation of CSF, white and gray matter volumes in hydrocephalic children using fuzzy clustering of MR images, Computerized Med. Imag. Graphics, vol. 18, no. 1, pp. 2534, 1994. [23] L. O. Hall, A. M. Bensaid, L. P. Clarke, R. P. Velthuizen, M. S. Silbiger, and J. C. Bezdek, A comparison of neural network and fuzzy clustering techniques in segmenting magnetic resonance images of the brain, IEEE Trans. Neural Networks, vol. 3, no. 5, pp. 672682, 1992. [24] M. Jungke, W. Von Seelen, G. Bielke, S. Meindl, G. Krone, M. Grigat, P. Higer, and P. Pfannenstiel, Information processing in nuclear magnetic resonance imaging, Magn. Reson. Imag., vol. 6, pp. 683693, 1988. [25] D. Y. Amamoto, R. Kasturi, and A. Mamourian, Tissue-type discrimination in magnetic resonance images, in Proc. 10th Int. Conf. Pattern Recogn., IEEE, 1990, pp. 603607. [26] M. ODonnell, J. C. Gore, and W. J. Adams, Toward an automated analysis system for nuclear magnetic resonance imagingII: Initial segmentation algorithm, Med. Phys., vol. 13, no. 3, pp. 293297, 1986. [27] D. A. Ortendahl, N. M. Hylton, L. Kaufman, and L. E. Crooks, Tissue characterization using intrinsic NMR parameters and a hierarchical processing algorithm, IEEE Trans. Nucl. Sci., vol. NS-32, no. 1, pp. 875879, 1985. [28] G. Cohen, N. C. Andreason, R. Alliger, S. Arndt, J. Kuan, W. T. C. Yuh, and J. Ehrhardt, Segmentation techniques for classication of brain tissue using magnetic resonance imaging, Psych. Res.: Neuroimag., vol. 45, pp. 3351, 1992. [29] T. L. Jernigan, G. A. Press, and J. R. Hesselink, Methods for measuring brain morphologic features on magnetic resonance images, Arch. Neurol., vol. 47, pp. 2732, 1990. [30] M. I. Kohn, N. K. Tanna, G. T. Herman, S. M. Resnick, P. D. Mozley, R. E. Gur, A. Alavi, R. A. Zimmerman, and R. C. Gur, Analysis of brain and cerebrospinal uid volumes with MR imagePart 1: Methods, reliability, and validation, Radiol., vol. 178, pp. 115122, 1991. [31] L. P. Clarke, R. P. Velthuizen, S. Phuphanich, J. D. Schellenberg, J. A. Arrington, and M. Silbiger, MRI: Stability of three supervised segmentation techniques, Magn. Reson. Imag., vol. 11, pp. 95106, 1993. [32] H. E. Cline, W. E. Lorensen, R. Kikinis, and F. Jolesz, Threedimensional segmentation of MR images of head using probability and

[33] [34] [35] [36]

[37] [38] [39] [40] [41] [42]

[43] [44] [45] [46] [47] [48] [49]

[50] [51] [52] [53]

[54] [55] [56] [57]

connectivity, J. Comput. Assist. Tomogr., vol. 14, no. 6, pp. 10371045, 1990. G. Gerig, J. Martin, R. Kikinis, O. Kubler, M. Shenton, and F. A. Jolesz, Unsupervised tissue type segmentation of 3-D dual-echo MR head data, Image and Vision Computing, vol. 10, no. 6, pp. 349360, 1992. M. W. Vannier, C. M. Speidel, and D. L. Rickman, Magnetic resonance imaging multispectral tissue classication, News Physiolog. Sci., vol. 3, pp. 148154, 1988. M. W. Vannier, R. L. Buttereld, D. Jordan, W. A. Murphy, R. G. Levitt, and M. Gado, Multispectral analysis of magnetic resonance images, Radiol., vol. 154, pp. 221224, 1985. Z. Liang, R. J. Jaszczak, and R. E. Coleman, Parameter estimation of nite mixtures using the EM algorithm and information criteria with application to medical image processing, IEEE Trans. Nucl. Sci., vol. 39, no. 4, pp. 11261133, 1992. Z. Liang, Tissue classication and segmentation of MR images, IEEE Eng. Med. Biol. Mag., pp. 8185, 1993. Z. Liang, J. R. MacFall, and D. P. Harrington, Parameter estimation and tissue segmentation from multispectral MR images, IEEE Trans Med. Imag., vol. 13, no. 3, pp. 339348, 1994. H. S. Choi, D. R. Haynor, and Y. Kim, Partial volume tissue classication of multichannel magnetic resonance imagesA mixed model, IEEE Trans Med. Imag., vol. 10, no. 3, pp. 395407, 1991. T. Taxt and A. Lundervold, Multispectral analysis of the brain using magnetic resonance imaging, IEEE Trans Med. Imag., vol. 13, no. 3, pp. 470481, 1994. A. Lundervold and G. Storvik, Segmentation of brain parenchyma and cerebrospinal uid in multispectral magnetic resonance images, IEEE Trans Med. Imag., vol. 14, no. 2, pp. 339348, 1995. E. A. Ashton, M. J. Berg, K. J. Parker, J. Weisberg, C. W. Chen, and L. Ketonen, Segmentation and feature extraction techniques with applications to MRI head studies, Magn. Reson. Med., vol. 33, pp. 670677, 1995. S. P. Raya, Low-level segmentation of 3-D magnetic resonance brain imagesA rule based system, IEEE Trans Med. Imag., vol. 9, no. 3, pp. 327337, 1990. H. Suzuki and J. Toriwaki, Automatic segmentation of head MRI images by knowledge guided thresholding, Computerized Med. Imag. Graphics, vol. 15, no. 4, pp. 233240, 1991. D. N. Kennedy, P. A. Filipek, and V. S. Caviness, Anatomic segmentation and volumetric calculations in nuclear magnetic resonance imaging, IEEE Trans Med. Imag., vol. 8, No. 1, pp. 17, 1989. G. Mittelhauber and F. Kruggel, Fast segmentation of brain magnetic resonance tomograms, in Conf. Comput. Vision, Virtual Reality, and Robotics, in Medicine, Apr. 1995, pp. 351357. W. C. Lin, E. C. Tsao, and C. T. Chen, Constraint satisfaction neural networks for image segmentation, Pattern Recogn., vol. 25, no. 7, pp. 679693, 1992. P. Santago and H. D. Gage, Quantication of MR brain images by mixture density and partial volume modeling, IEEE Trans. Med. Imag., vol. 12, no. 3, pp. 566574, 1993. C. DeCarli, J. Maisog, D. G. M. Murphy, D. Teichberg, S. I. Rapoport, and B. Horwitz, Method for quantication of brain, ventricular and subarachnoid CSF volumes from MR images, J. Comput. Assist. Tomogr., vol. 16, no. 2, pp. 274284, 1992. D. L. Toulson and J. F. Boyce, Segmentation of MR images using neural nets, Image Vision, Computing, vol. 5, pp. 324328, 1992. M. M. Chang, A. M. Teklap, and M. I. Sezan, Bayesian segmentation of MR images using 3-D Gibbsian priors, in Image Video Processing, SPIE Proc., 1993, vol. 1903, pp. 122133. J. C. Bezdek, L. O. Hall, and L. P. Clarke, Review of MR image segmentation techniques using pattern recognition, Med. Phys., vol. 20, no. 4, pp. 10331048, 1993. L. P. Clarke, R. P. Veltuizen, M. A. Camacho, J. J. Heine, M. Vaidyanathan, L. O. Hall, R. W. Thatcher, and M. L. Silbiger, MRI segmentation: Methods and applications, Magn. Reson. Imag., vol. 13, no. 3, pp. 343368, 1995. M. Morrison and Y. Attikiouzel, An introduction to the segmentation of magnetic resonance medical images, Australian Comput. J., vol. 26, no. 3, pp. 9098, 1994. G. Gerig, O. Kubler, R. Kikinis, and F. A. Jolesz, Nonlinear anisotropic ltering of MRI data, IEEE Trans Med. Imag., vol. 11, no. 2, pp. 221232, 1992. K. M. Ludeke, P. Roschmann, and R. Tischler, Susceptibility artifacts in NMR imaging, Magn. Reson. Imag., vol. 3, pp. 329343, 1985. B. R. Condon, J. Patterson, D. Wyper, A. Jenkins, and D. M. Hardley, Image nonuniformity in magnetic resonance imaging: Its magnitude and methods for its correction, Br. J. Radiol., vol. 60, pp. 8387, 1987.

186

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 16, NO. 2, APRIL 1997

[58] D. A. G. Wicks, G. J. Barker, and P. S. Tofts, Correction of intensity nonuniformity in MR images of any orientation, Magn. Reson. Imag., vol. 11, pp. 183196, 1993. [59] B. M. Dawant, A. P. Zijdenbos, and R. A. Margolin, Correction of intensity variations in MR images for computer-aided tissue classication, IEEE Trans Med. Imag., vol. 12, no. 4, pp. 770771, 1993. [60] M. Tincher, C. R. Meyer, R. Gupta, and D. M. Williams, Polynomial modeling and reduction of RF body coil spatial inhomogeneity in MRI, IEEE Trans Med. Imag., vol. 12, no. 2, pp. 361365, 1993. [61] H. Chang and M. Fitzpatrick, A technique for accurate magnetic resonance imaging in the presence of eld inhomogeneities, IEEE Trans Med. Imag., vol. 11, no. 3, pp. 319329, 1992. [62] M. C. de Oliveira and R. I. Kitney, Texture analysis for discrimination of tissues in MRI data, in Proc. Computers in Cardiology, IEEE, 1991, pp. 481484. [63] R. M. Haralick, K. Shanmugam, and I. Dinstein, Textural features for image classication, IEEE Trans. Syst., Man, Cybern., vol. SMC-3, no. 6, pp. 610621, 1973. [64] H. A. Koenig, R. Bachus, and E. R. Reinhardt, Pattern recognition for tissue characterization in magnetic resonance imaging, Health Care Instrum., pp. 183187, 1986. [65] L. Kjaer, C. Thomsen, F. Gjerris, B. Mosdal, and O. Henriksen, Tissue characterization of intracranial tumors by MRI imaging, Acta Radiologica, vol. 32, pp. 498504, 1989. [66] T. J. Hyman, R. J. Kurland, G. C. Levy, and J. D. Shoop, Characterization of normal brain tissue using seven calculated MRI parameters and a statistical analysis system, Magn. Reson. Med., vol. 11, pp. 2234, 1989. [67] R. A. Redner and H. F. Walker, Mixture densities, maximum likelihood, and the EM algorithm, SIAM Rev., vol. 26, no. 2, pp. 195239, 1984. [68] D. M. Titterington, A. F. M. Smith, and U. E. Makov, Statistical Analysis of Finite Mixture Distributions. New York: Wiley, 1985. [69] J. Besag, Spatial interaction and the statistical analysis of lattice systems, J. Roy. Statist. Soc. B, vol. 36, no. 2, pp. 192225, 1974. [70] M. Hassner and J. Sklansky, The use of Markov random elds as models of texture, Comput. Graphics, Image Processing, vol. 12, pp. 357374, 1980. [71] F. R. Hansen and H. Elliot, Image segmentation using simple Markov eld models, Comput. Graphics, Image Processing, vol. 20, pp. 101132, 1982. [72] S. Geman, and D. Geman, Stochastic relaxation, Gibbs distribution, and the Bayesian restoration of images, IEEE Trans. Pattern Anal. Machine Intell., vol. PAMI-6, pp. 721741, 1984. [73] F. S. Cohen and D. Cooper, Simple parallel hierarchical and relaxation algorithms for segmenting noncausal Markovian random elds, IEEE Trans. Pattern Anal. Machine Intell., vol. PAMI-9, no. 2, pp. 195219, 1987. [74] R. C. Dubes and A. K. Jain, Random eld models in image analysis, J. Appl. Statistics, vol. 16, no. 2, pp. 131164, 1989.

[75] B. S. Manjunath and R. Chelappa, Unsupervised texture segmentation using Markov random eld models, IEEE Trans. Pattern Anal. Machine Intell., vol. 13, no. 5, pp. 478482, 1991. [76] R. Leahy, T. Hebert, and R. Lee, Applications of Markov random elds in medical imaging, Inform. Processing Med. Imag., pp. 114, 1991. [77] R. O. Duda and P. E. Hart, Pattern Classication and Scene Analysis. New York: Wiley, 1973. [78] A. P. Dempster, L. N. Laird, and D. B. Rubin, Maximum likelihood from incomplete data via the EM algorithm, J. Roy. Statist. Soc., vol. 38-B, pp. 138, 1977. [79] N. Karssemeijer, A statistical method for automatic labeling of tissues in medical images, Machine Vision, Applicat., vol. 3, pp. 7586, 1990. [80] J. Besag, On the statistical analysis of dirty pictures, J. Roy. Statist. Soc. B, vol. 48, no. 3, pp. 259302, 1986. [81] J. Marroquin, S. Mitter, and T. Poggio, Probabilistic solution of illposed problems in computational vision, J. Amer. Statistical Assoc., vol. 82, no. 397, pp. 7689, 1987. [82] T. N. Pappas, An adaptive clustering algorithm for image segmentation, IEEE Trans. Signal Processing, vol. 40, no. 4, pp. 901914, 1992. [83] H. Derin and H. Elliott, Modeling and segmentation of noisy and textured images using Gibbs random elds, IEEE Trans. Pattern Anal. Machine Intell., vol. PAMI-9, no. 1, pp. 3955, 1987. [84] S. Lakshmanan and H. Derin, Simultaneous parameter estimation and segmentation of Gibbs random elds using simulated annealing, IEEE Trans. Pattern Anal. Machine Intell., vol. 11, no. 8, pp. 799813, 1989. [85] S. Kirkpatrck, C. D. Gelatt, and M. P. Vecchi, Optimization by simulated annealing, Sci., vol. 220, No. 4598, pp. 671680, 1983. [86] S. L. Sclove, Application of the conditional population mixer model to image segmentation, IEEE Trans. Pattern Anal. Machine Intell., vol. PAMI-5, no. 4, pp. 428433, 1983. [87] M. R. Anderberg, Cluster Analysis for Applications. New York: Academic, 1973. [88] J. W. Snell, M. Merickel, J. Ortega, J. Goble, J. Brookeman, and N. Kassell, Boundary estimation of complex objects using hierarchical active surface templates, Pattern Recogn., vol. 28, no. 10, pp. 15991609, 1995. [89] J. N. Giedd, J. W. Snell, N. Lange, J. C. Rajapakse, P. Kozuch, B. J. Casey, D. Kaysen, A. C. King, S. D. Hamburger, and J. L. Rapoport, Quantitative magnetic resonance imaging of human brain development: Age 518, Cerebral Cortex, vol. 6, no. 4, pp. 110, 1996. [90] J. W. Bulte, T. Douglas, S. Mann, R. B. Grankel, B. M. Moskowitz, R. A. Brooks, C. D. Baumgarner, J. Vymazal, M. Strub, and J. A. Frank, Magnetoferritin: Characterization of a novel superparamagnetic MR contrast agent, J. Magn. Reson. Imag., vol. 4, pp. 497505, 1994. [91] J. C. Rajapakse, C. DeCarli, A. Mclaughlin, J. N. Giedd, A. L. Krain, S. D. Hamburger, and J. L. Rapoport, Cerebral magnetic resonance image segmentation using data fusion, J. Comput. Assist. Tomogr., vol. 20, no. 2, pp. 206218, 1996.

You might also like