Segmentation of corpus callosum using diffusion tensor imaging: validation in patients with glioblastoma
 4.9k Downloads
 6 Citations
Abstract
Background
This paper presents a threedimensional (3D) method for segmenting corpus callosum in normal subjects and brain cancer patients with glioblastoma.
Methods
Nineteen patients with histologically confirmed treatment naïve glioblastoma and eleven normal control subjects underwent DTI on a 3T scanner. Based on the information inherent in diffusion tensors, a similarity measure was proposed and used in the proposed algorithm. In this algorithm, diffusion pattern of corpus callosum was used as prior information. Subsequently, corpus callosum was automatically divided into Witelson subdivisions. We simulated the potential rotation of corpus callosum under tumor pressure and studied the reproducibility of the proposed segmentation method in such cases.
Results
Dice coefficients, estimated to compare automatic and manual segmentation results for Witelson subdivisions, ranged from 94% to 98% for control subjects and from 81% to 95% for tumor patients, illustrating closeness of automatic and manual segmentations. Studying the effect of corpus callosum rotation by different Euler angles showed that although segmentation results were more sensitive to azimuth and elevation than skew, rotations caused by brain tumors do not have major effects on the segmentation results.
Conclusions
The proposed method and similarity measure segment corpus callosum by propagating a hypersurface inside the structure (resulting in high sensitivity), without penetrating into neighboring fiber bundles (resulting in high specificity).
Keywords
Corpus callosum Fiber bundle segmentation Levelset Glioblastoma Diffusion tensor imagingBackground
Corpus callosum is the largest interhemispheric fiber bundle in the human brain [1, 2]. Most of the fibers interconnect homologue cortical areas in roughly mirror image sites but a large number of the fibers have heterotypic connections ending in asymmetrical areas [3]. Previous studies have mainly investigated effects of various pathologies on the corpus callosum [4, 5, 6, 7]. However, a fully automated, fast, and accurate method for segmenting corpus callosum without penetrating into irrelevant neighboring structures, using data acquired in routine clinical protocols, is still lacking.
Previously, image processing methods have been proposed for segmenting corpus callosum in anatomical magnetic resonance images (MRI) [8, 9, 10]. These methods rely on intensity information of twodimensional images and their results may need pruning. Recently, attention has been oriented towards diffusion tensor imaging (DTI) to segment white matter tracts of the brain [11, 12]. Although the tensor model fails to describe higher order anisotropies in heterogeneous areas where more than one fiber population exists, it is practically useful for extracting major white matter tracts, particularly the ones with predominant diffusivity pattern such as corpus callosum. When using DTI data, the fiber bundles can be extracted by: a) clustering of fibers resulting from tractography into fiber bundles [13, 14, 15, 16, 17, 18, 19]; or b) segmenting fiber bundles via hypersurface propagation based on local properties of diffusion tensor, diffusion signal, or orientation distribution function (ODF) [20, 21, 22, 23, 24, 25, 26, 27]. Since clustering methods rely on the tractography results, they do not work properly if the tractography results are inaccurate. On the other hand, segmentation methods based on hypersurface propagation do not use the tractography results and are thus more robust to noise.
In a regionbased segmentation framework, a similarity measure between successive tensors is typically used. Some of the hypersurface propagating methods in the literature concentrated on scalar quantities derived from the tensor data which do not reflect complete tensor information [20]. Other methods benefit from the entire information contained in the DTI data [21, 22, 23, 24, 25, 26, 27, 28, 29, 30]. Wang and Vemuri [22] proposed a statistical levelset segmentation method. However, the tensors derived in this framework are not necessarily positive semidefinite, leading to inappropriate results especially when consecutive tensors are much different. Metrics like KullbackLeibler divergence and Jdivergence [23, 24] have also been proposed. One of the most promising methods is introduced by Jonasson et al. [25]. They defined a new similarity measure called normalized tensor scalar product (NTSP). Comparing NTSP with other similarity measures, they demonstrated superiority of their proposed measure. To segment brain structures like thalamic nuclei, they modified their framework to favor the propagation of multiple hyper surfaces without overlapping [26]. Lenglet et al. [27] defined a dissimilarity measure and statistics between tensors based on the Riemannian distances. Although improving the segmentation results, this approach is computationally expensive. Defining a LogEuclidean distance, another metric was defined by Arsigny et al. [28] which has lower computational burden. Weldeselassie and Hamarneh [29] used their proposed similarity measure in an energy minimization framework. Awate et al. [30] used the similarity measure in a Markov random field framework.
In terms of quantitative evaluation of diffusion parameters, previous studies have compared DTIbased indices in normal appearing white matter and corpus callosum in multiple sclerosis [4], stroke [5], schizophrenia [6], and Huntington's [7] and also studied the DTI methods to assess corpus callosum regions across the human lifespan [31]. For segmenting corpus callosum and its subdivisions in these studies, however, twodimensional (2D) methods were applied and DTIbased indices compared in the midsagittal plane. However, without recruiting a three dimensional (3D) method to segment the whole corpus callosum and its subdivisions, the extracted quantities may be inaccurate.
Since the tensor model is not capable of describing heterogeneous diffusion behavior in the crossing fiber bundles, some studies used High Angular Resolution Diffusion Imaging (HARDI) data to segment specific bundles [32, 33, 34, 35, 36, 37, 38]. However, the HARDI data is not widely acquired in clinical centers and hence, the tensorbased methods are still of more practical use in clinical research.
In this manuscript, we present a 3D method to segment corpus callosum. Based on tensor and anisotropy values of neighboring voxels, a similarity measure is proposed and used as a speed function in the proposed levelset method. In this method, the principal diffusion direction (PDD) and prior information about the diffusivity pattern in corpus callosum are used to avoid inclusion of neighboring fiber bundles. Then, the Witelson subdivisions of corpus callosum are automatically identified [39]. The idea of using diffusivity pattern in corpus callosum has been used by Lee et al. [40]. However, they performed a 2D segmentation on the midsagittal plane. Moreover, since their method uses the leftright component of PDD to delineate corpus callosum boundaries, it is not applicable in more lateral sagittal planes, where corpus callosum connects to minor and major forceps and considerable anteriorposterior component of PDD exists.
The proposed segmentation method can be identically used for segmenting corpus callosum of control subjects as well as patients with glioblastoma. However, since we use the geometric information of the diffusivity pattern in corpus callosum and the glioblastoma tumor may change the original shape and diffusivity pattern of corpus callosum, we validate the reproducibility of the proposed method through realistic simulations of corpus callosum rotations under glioblastoma tumor pressure. We study the effects of rotation by different Euler angles (azimuth, elevation and skew) quantitatively and demonstrate that even in extreme cases with large rotations, brain tumors do not have major effects on the segmentation results generated by our proposed method. We apply the method to the DTI data of normal subjects and brain cancer patients with glioblastoma and show its superiority to some of the previously published methods in the literature.
Methods
Patient population
This study is approved by the institutional review board and is compliant with the Health Insurance Portability and Accountability Act (HIPAA). Between February 2006 and December 2008, 19 patients (8 males, 11 females; mean age 60.7 years) with treatment naïve glioblastoma underwent MRI with DTI using a 3 T scanner at our institution. Based on anatomical MRI findings, the patients were divided into two groups: Group 1 including patients with tumors not infiltrating corpus callosum (n = 12); and Group 2 including patients with tumors infiltrating corpus callosum (n = 7). For the control group, 11 patients (5 males, 6 females; mean age 48 years) who underwent brain MRI for nonspecific headache or single idiopathic seizure with normal MRI were included.
Magnetic resonance imaging protocol
All the patients underwent both conventional MRI and DTI on a 3 T scanner (Excite HD, GE Medical Systems, Milwaukee, WI) using an 8channel head coil. Diffusion weighted images (DWIs) were acquired in 25 diffusion gradient directions. The reconstructed DWIs have intraslice resolution of 256 × 256 with voxel size of 0.98 × 0.98 × 2.5 mm. To have the same step size in each direction for the front propagation in the levelset method and to avoid extensive computation, we interpolated the data into 128 × 128 grid with cubical, 1.9 × 1.9 × 1.9 mm voxels.
Levelset approach
We use a levelset method that takes into account several properties when formulating the segmentation problem [41, 42]. The method smoothes the propagating hypersurface automatically and leads to a regularized segmentation. Among the advantages of the method, its property of generalizing from 2D to 3D and higher dimensions and automatic splitting and merging of the surfaces are notable. In addition, it simplifies calculation of geometric quantities needed in the proposed method such as normal to surface and curvature.
where r ∈ ℜ^{n} is the state space, φ: ℜ^{n} × ℜ → ℜ is the levelset function, D_{t} φ is the partial derivative of φ with respect to the time variable t, ∇φ = D_{r} φ is the gradient of φ with respect to the state space variables, and F(r, t) is the speed in the direction normal to the surface, extracted from the spherical harmonic coefficients of the neighboring voxels. The sign ∥·∥ stands for the magnitude operator. The curvature κ(r, t) is used to fulfill the smoothness constraint. To calculate the first order spatial partial derivative ∇φ(r, t), we use the 5th order of the upwind method [41].
Corpus callosum segmentation
Corpus callosum is a commissural fiber bundle with a specific diffusion pattern which can be coded as prior knowledge in the segmentation framework. Using this information, we prevent the hypersurface from propagating into adjacent white matter structures such as cingulum, tapetum, minor and major forceps, and tracts of the corona radiata. Although dissimilarity among tensors helps in this case, smooth and gradual transition in shape and direction of the DTI tensors from corpus callosum to minor and major forceps makes the segmentation difficult.
We define a similarity measure between every voxel on the propagating hypersurface and its neighbors in the propagation direction, based on tensor and anisotropy values. This similarity measure is used as the speed term F(r, t) in Equation (1). The hypersurface propagation in each step depends only on the speed function of the boundary voxels. This speed function term moves the hypersurface to fill the whole fiber bundle, while the regularizing curvature term in Equation (1) is in charge of smoothing corpus callosum without changing its real structure.
Using the fact that the diffusivity in corpus callosum is perpendicular to the midsagittal plane of the brain, we consider a threshold (PDD_{ x } Threshold) for the xcomponent (leftright component) of the PDD (PDD_{ x }(r)) for propagating the front only inside the corpus callosum. The xcomponent of the PDD is large throughout the structure body. However, corpus callosum fibers project to the cortical area at its genu and splenium, where the xcomponent of PDD is not large anymore. Fortunately, these fibers are considered as different fiber tracts of minor and major forceps, and most of studies investigating the different diffusion indices within the corpus callosum, exclude the minor and major forceps from that.
In the proposed algorithm, we consider two more thresholds on collinearity of the PDD vectors (Collinearity_Threshold) and similarity of fractional anisotropy (FA) values (FA_Threshold) in the neighboring voxels.
 1.
Select the initial seeds in corpus callosum in midsagittal plane manually.
 2.
Initiate the hypersurface as the congregation of small spheres around the seed points.
 3.Do until convergence
 For each point r on the hypersurface at step t:
 a)
Calculate the normal direction to the surface.
 b)
Calculate the 26neighborhood and keep the neighbors n_{ r } for r, which are collinear with the normal, with respect to the r.
 c)
If PDD(r). PDD(n_{ r }) > Collinearity_Threshold & FA (n_{ r }) > FA_Threshold & PDD_{x} (r) > PDD_{ x }_Threshold
 a)

Threshold F(r, t) with F _ Threshold to diminish the effect of negligible speeds.

Use the resultant speed in the levelset framework.
$\text{Then}\phantom{\rule{1em}{0ex}}F\left(r,t\right)=\sum _{{n}_{r}}FA\left(r\right).FA\left({n}_{r}\right).\frac{tr\left[D{\left(r\right)}^{*}D\left({n}_{r}\right)\right]}{tr{\left[D\left(r\right)\right]}^{*}tr\left[D\left({n}_{r}\right)\right]}$(2) 
 4.
Extract the zero levelset as the segmented corpus callosum.
In Equation (2), tr(.) is the matrix trace and D(r) is the tensor at point r.
After segmenting corpus callosum, Witelson subdivisions of corpus callosum are automatically extracted [39]. First, the critical point between genu and rostrum of corpus callosum is calculated, where the curvature of the structure boundary in midsagittal plane changes. Then, the segmented corpus callosum is automatically subdivided into Witelson subdivisions in the midsagittal plane: rostrum, genu, rostral body, anterior midbody, posterior midbody, isthmus and splenium [39]. Moreover, the user can visualize and confirm the calculated midsagittal plane and the critical joining point between genu and rostrum. The critical point can be selected manually if close supervision is preferred or needed.
Results
Segmentation using proposed method
We implemented the proposed algorithm in MATLB R2008a using a PC with Intel^{®} Core™ 2Duo CPU (E8400@ 3.00 GHz, 3.00 GHz) and 4 GByte RAM and 64 bit VISTA operating system.
where S_{a} and S_{ r } are the automatic and manual (reference) segmentation results, respectively, and N is the number of voxels in each bundle. Here, N(S_{a} ∩ S_{y})is the number of the TruePositives. The Dice correctness measure is appropriately bounded, normalized, wellunderstood, and applied widely in evaluating segmentation methods. Two of the coauthors (a clinical expert and a technical expert) sat down together and carried out the manual segmentation which is considered as the reference segmentation results.
Comparison of the proposed method with the Jonasson's method
Witelson Subdivisions  Proposed Method  Jonasson's Method 

Rostrum  94.41  75.15 
Genu  98.25  91.24 
Rostral Body  97.27  83.10 
Anterior MidBody  95.44  79.33 
Posterior MidBody  96.07  71.54 
Isthmus  94.24  81.46 
Splenium  97.29  80.25 
Average Dice measures for different Witelson subdivisions of the corpus callosum
Witelson Subdivisions  Control Subjects  Group 1 Patients  Group 2 Patients 

Rostrum  94.41  94.04  90.72 
Genu  98.25  94.99  92.61 
Rostral Body  97.27  94.05  84.73 
Anterior MidBody  95.44  91.45  81.58 
Posterior MidBody  96.07  94.1  83.29 
Isthmus  94.24  94.57  85.09 
Splenium  97.29  90.16  89.56 
Rotational effect of tumor on proposed method
We have used the geometric information of the diffusivity pattern in corpus callosum to prevent the front from penetrating into the neighboring structures. However, a tumor may change the shape and diffusivity pattern of corpus callosum from its original shape and diffusivity pattern.
where (x, y, z) is the coordinates for a voxel within a 5voxel neighborhood of corpus callosum, while (x', y'z') is the rotated voxel coordinates.
Means and standard deviations of the Dice measures in 11 normal subjects where corpus callosum subdivisions are rotated under different Euler angles
Witelson Subdivisions of Corpus Callosum  Rostrum  Genu  Rostral Body  Anterior MidBody  Posterior MidBody  Isthmus  Splenium  

Azimuth Rotation Angle  5  94.51 ± 1.15  96.46 ± 1.43  99.07 ± 0.84  98.48 ± 0.77  99.28 ± 0.56  96.84 ± 0.91  96.23 ± 0.93 
10  91.44 ± 0.99  95.48 ± 0.98  95.02 ± 1.13  95.45 ± 0.99  97.36 ± 1.03  96.27 ± 0.93  94.11 ± 1.01  
15  91.00 ± 1.47  90.77 ± 1.15  94.17 ± 0.95  94.47 ± 0.96  94.97 ± 0.70  92.84 ± 1.01  90.55 ± 0.97  
20  88.29 ± 1.24  89.33 ± 1.07  89.65 ± 0.75  90.61 ± 1.23  88.86 ± 0.86  89.18 ± 1.23  88.40 ± 1.06  
25  84.24 ± 1.03  88.47 ± 1.12  89.16 ± 0.86  87.32 ± 1.03  88.83 ± 1.06  89.01 ± 1.49  84.69 ± 1.20  
30  80.16 ± 1.70  82.96 ± 1.23  85.09 ± 1.21  85.45 ± 1.13  85.36 ± 0.98  83.07 ± 1.11  81.23 ± 0.97  
Elevation Rotation Angle  5  93.64 ± 0.92  96.29 ± 1.28  97.75 ± 0.97  96.91 ± 1.02  98.65 ± 1.16  96.29 ± 0.72  95.83 ± 1.09 
10  91.27 ± 0.88  96.24 ± 1.08  97.74 ± 1.00  95.46 ± 1.06  94.14 ± 1.09  91.56 ± 1.17  91.96 ± 1.41  
15  87.75 ± 0.86  95.28 ± 1.15  93.86 ± 0.89  93.81 ± 0.97  94.95 ± 0.83  89.30 ± 0.91  91.08 ± 1.14  
20  85.37 ± 1.01  92.58 ± 0.77  90.59 ± 0.89  89.98 ± 1.29  92.51 ± 1.05  87.36 ± 1.32  86.29 ± 0.88  
25  83.47 ± 0.90  86.67 ± 0.77  86.25 ± 1.32  89.27 ± 1.17  87.66 ± 0.99  84.88 ± 1.27  84.48 ± 1.10  
30  79.39 ± 1.29  81.60 ± 1.10  85.00 ± 1.10  84.35 ± 1.01  85.22 ± 1.03  83.41 ± 0.88  81.58 ± 0.89  
Skew Rotation Angle  5  97.96 ± 1.09  98.75 ± 0.63  99.09 ± 0.68  98.00 ± 1.07  98.73 ± 1.00  96.62 ± 0.73  98.85 ± 0.59 
10  95.29 ± 1.12  96.38 ± 0.80  94.72 ± 1.04  94.46 ± 1.45  94.80 ± 1.28  94.98 ± 1.00  94.21 ± 0.88  
15  91.88 ± 1.04  95.00 ± 1.34  92.56 ± 1.16  94.22 ± 0.97  92.73 ± 1.31  94.72 ± 1.44  93.66 ± 1.01  
20  89.79 ± 0.83  87.92 ± 1.18  85.83 ± 1.29  88.35 ± 1.16  86.99 ± 1.27  87.85 ± 0.98  89.28 ± 1.16  
25  87.55 ± 1.22  86.11 ± 1.21  85.73 ± 1.08  86.34 ± 1.28  86.22 ± 1.08  85.55 ± 1.27  87.35 ± 0.87  
30  84.61 ± 1.17  83.39 ± 0.89  84.58 ± 1.13  84.09 ± 1.03  85.26 ± 1.12  86.43 ± 1.16  85.00 ± 1.00 
The pvalue of Wilcoxon twosample tests between outer (rostrum, genu, and splenium) and inner (rostral body, anterior and posterior midbody, and isthmus) subdivisions of corpus callosum
Azimuth Rotation Angle  

5  10  15  20  25  30  
P value  8.2E05  8.1E05  8.1E05  7.1E03  8.2E05  8.2E05 
Elevation Rotation Angle  
5  10  15  20  25  30  
P value  8.2E05  3.9E04  1.4E04  8.2E05  8.0E05  8.2E05 
Skew Rotation Angle  
5  10  15  20  25  30  
P value  7.6E02  8.2E02  9.2E01  2.4E04  5.8E03  3.6E02 
Discussion
Novel aspects of proposed method
1. New similarity measure based on local diffusion characteristics
The proposed method is more accurate than the tensorbased method that uses the NTSP similarity measure [25]. Comparing with the manual segmentation by experts, we demonstrated accuracy of our proposed method even when the tumor infiltrates corpus callosum.
2. 3D segmentation of the entire corpus callosum
The proposed method works in 3D. Extraction of diffusion indices such as mean diffusivity and fractional anisotropy over the entire corpus callosum generates a reliable quantification of the structure that cannot be achieved by an analysis of the midsagittal plane only [4, 5, 6, 7].
3. Preventing from penetrating inside neighboring fiber bundles
An optimal set of parameters is chosen to segment the entire corpus callosum (with high sensitivity) without penetrating into adjacent fiber structures (with high specificity). We have used Collinearity_Threshold to prevent the front from penetrating into cingulum. The threshold for the xcomponent of the principal diffusion direction (PDD_{ x }_Threshold) is used to prevent the front from propagating into major and minor forceps. FA_Threshold is also important to prevent from penetrating inside tapetum (as in general, it has diffusivity patterns similar to corpus callosum in crossing areas with lower FA values). To segment the entire corpus callosum in patients with low FA values, one should choose FA_Threshold and F_Threshold quite small. However, one should be cautious that in this case, the front may propagate outside of white matter in some areas.
4. Automatically extracting Witelson subdivisions of corpus callosum
The proposed method defines Witelson subdivisions of corpus callosum automatically.
5. Applicability to glioblastoma tumor patients as well as normal subjects using the same set of parameters
The proposed method has successfully segmented corpus callosum and its subdivisions in diffusions MRI data of normal subjects and brain tumor patients.
6. Evaluating potential effects of tumor on segmentation of corpus callosum
Depending on size, shape, type and proximity to the corpus callosum, the glioblastoma tumor may cause rotation, shrinkage or more severe disruptions like tearing of the corpus callosum fibers. Amongst the mentioned effects, we evaluated the linear rotational effect of tumor on corpus callosum. With this simplification and without loss of generality, we showed that the change in diffusivity pattern due to a tumor does not change the segmentation accuracy dramatically. However, if the effect of tumor infiltration inside the corpus callosum is sever and changes the structure dramatically, it may be impossible to segment the structure entirely and accurately.
Selection of initial seeds
The proposed segmentation method is seedbased and needs the initial seed points for hypersurface to propagate. This requires the operators to define the seeds manually using anatomical landmarks. To automate the process, the initial seeds may be obtained from fiber atlases of control subjects [34, 44]. However, when segmenting fiber bundles of pathological cases, abnormalities like tumors may change the fiber bundles and thus conventional atlas registration methods may not be applicable. To solve this problem, Zacharaki et al. [45] proposed a method for transferring structural and functional information from neuroanatomical brain atlases into individual patient's data. Application of such methods in our case of segmenting corpus callosum in patients with tumors would be still in question.
Segmentation of fiber structures in HARDI
Since the tensor model is not capable of describing heterogeneous diffusion behavior in crossing fiber bundles, some studies segmented the desired fiber bundles using HARDI data. Generalizing the levelset method presented in [22] to the HARDI data, Descoteaux and Deriche [35] applied a regionbased statistical surface evolution to the image of the ODFs to find coherent white matter fiber bundles. This is equivalent to the maximization of a posteriori probability which obtains the desired segmentation for the observed ODFs and presumes Gaussian distributions in different partitions of the Qball images. Although their method appropriately propagates through the regions of fiber crossings, it propagates the hypersurface inside all of the crossing fiber bundles and segments them as a whole, not individually. They applied their method to extract corpus callosum and tracts of coronaradiata. However, their method is sensitive to initial seeds and suffers from limited anatomical knowledge of the operator who defines the seeds. In another study, a kmeans algorithm has been employed to find the clusters using Euclidean distance as dissimilarity measure [36].
In a different study, a PositionOrientation Space (POS) is introduced by combining the geometric space with the spherical ODF space from the HARDI data, where two crossing fiber populations with different orientations in the spatial domain are resolved by applying a front propagation method in the levelset framework in a fivedimensional space [32]. Hagmann et al. [36] performed fiber bundle segmentation in POS based on the Markov Random Fields (MRF). In a similar study, McGraw et al. [37] proposed a mixture of von MisesFisher distributions to model the ODF again in MRF. Assuming that the spatial relationships are modeled by the MRF, this method estimates a hidden random field of fiber bundles from the observed ODF profiles using a Maximum a Posteriori (MAP) formulation. However, dimensional reinforcement of the problem causes disadvantages such as increasing the computational cost.
Using Spherical Harmonic Coefficients (SHC) as features of functions on the sphere [38], a method has been proposed for fiber bundle segmentation [33]. However, without masking the speed term with a measure of anisotropy (such as FA) that has low values in the crossing areas, the growing hypersurface may penetrate into irrelevant fiber bundles that have common areas with the bundle of interest. In another work [34], the authors proposed an atlasbased method introducing a novel similarity based on PDD's and spherical harmonic coefficients. Integrating PDD's into the framework, along with a proper PDDselection algorithm, leads to the segmentation of most of important fiber bundles in the brain without penetration into irrelevant fiber bundles that have common areas with the bundle of interest. Using an atlas [44] to find the initial seeds for the fiber bundles, the proposed method overcomes limitations of the semiautomatic methods that suffer from limited anatomical knowledge and subjectivity of the operator who defines the seed voxels. Note that since the number of diffusion measurements in the tensor data is not adequate for fitting SHC and estimating more than one PDD, we used the same methods originally applied on the HARDI data.
Although the success of recent studies in segmenting fiber bundles in the HARDI data is promising, such protocol in not being widely used in clinical centers. On the other hand, diffusion tensor imaging is clinically feasible and thus the tensorbased methods are of more interest.
Conclusion
In this paper, we have proposed a 3D method based on a new DTI similarity measure to segment corpus callosum and determine its subdivisions. The method propagates a hypersurface within corpus callosum without penetration into the neighboring fiber bundles. Segmentation of corpus callosum, the largest commissural fiber bundle in the brain, makes it possible to quantify various diffusion characteristics in its subdivisions, opening new perspectives for monitoring disease evolution or prognosis.
Ethics statement
MRI and other data of the tumor patients originally acquired for patient care are retrospectively and anonymously used in this research to train, test, and evaluate the proposed methods. This usage was reviewed and approved by the Institutional Review Board (IRB) committee of Henry Ford Hospital, Detroit, Michigan, USA. Details of the data and results are described in the manuscript.
Notes
Supplementary material
References
 1.de Lacoste MC, Kirkpatrick JB, Ross ED: Topography of the human corpus callosum. J Neuropathol Exp Neurol. 1985, 44: 578591. 10.1097/0000507219851100000004.CrossRefPubMedGoogle Scholar
 2.Clarke S, Miklossy J: Occipital cortex in man: organization of callosal connections, related myelo and cytoarchitecture, and putative boundaries of functional visual areas. J Comp Neurol. 1990, 298: 188214. 10.1002/cne.902980205.CrossRefPubMedGoogle Scholar
 3.Dorion A, Chantôme M, Hasboun D, et al: Hemispheric asymmetry and corpus callosum morphometry: a magnetic resonance imaging study. Neuroscience Research. 2000, 36: 913. 10.1016/S01680102(99)001029.CrossRefPubMedGoogle Scholar
 4.Hasan K, Gupta R, Santos R, Wolinsky J, Narayana P: Diffusion tensor fractional anisotropy of the normalappearing seven segments of the corpus callosum in healthy adults and relapsingremitting multiple sclerosis patients. Journal of Magnetic Resonance Imaging. 2005, 21: 735743. 10.1002/jmri.20296.CrossRefPubMedGoogle Scholar
 5.Gupta R, Saksena S, Hasan K, et al: Focal Wallerian degeneration of the corpus callosum in large middle cerebral artery stroke: serial diffusion tensor imaging. Journal of Magnetic Resonance Imaging. 2006, 24: 549555.CrossRefPubMedGoogle Scholar
 6.Kubicki M, Styner M, Bouix S, et al: Reduced interhemispheric connectivity in schizophreniatractography based segmentation of the corpus callosum. Schizophrenia research. 2008, 106: 125131. 10.1016/j.schres.2008.08.027.CrossRefPubMedPubMedCentralGoogle Scholar
 7.Rosas HD, Lee SY, Bender AC, Zaleta AK, Vangel M, Yu P, Fischl B, Pappu V, Onorato C, Cha JH, Salat DH, Hersch SM: Altered white matter microstructure in the corpus callosum in Huntington's disease: Implications for cortical. Neuroimage. 2010, 49: 29953004. 10.1016/j.neuroimage.2009.10.015.CrossRefPubMedGoogle Scholar
 8.Lundervold A, Duta N: Taxt T. Jain A: Modelguided segmentation of corpus callosum in MR images. Citeseer. 1999, 231237.Google Scholar
 9.Lee C, Huh S, Ketter T, Unser M: Automated segmentation of the corpus callosum in midsagittal brain magnetic resonance images. Optical Engineering. 2000, 39: 924935. 10.1117/1.602449.CrossRefGoogle Scholar
 10.Brejl M, Sonka M: Object localization and border detection criteria design in edgebased image segmentation: automated learning from examples. IEEE Transactions on Medical Imaging. 2000, 19: 973985. 10.1109/42.887613.CrossRefPubMedGoogle Scholar
 11.Basser P, Mattiello J, Le Bihan D: MR diffusion tensor spectroscopy and imaging. Biophys. 1994, 66: 259267. 10.1016/S00063495(94)807751.Google Scholar
 12.Basser P, Mattiello J, Le Bihan D: Estimation of the effective selfdiffusion tensor from the NMR spin echo. J Magn Reson. 1994, 103: 247254. 10.1006/jmrb.1994.1037.CrossRefGoogle Scholar
 13.Brun A, Knutsson H, Park H, Shenton M, Westin C: Clustering fiber traces using normalized cuts. MICCAI LNCS. 2004, 3216: 368375.Google Scholar
 14.Maddah M, Grimson W, Warfield S, Wells W: A unified framework for clustering quantitative analysis of white matter fiber tracts. Med Imag Analysis. 2008, 12: 191202. 10.1016/j.media.2007.10.003.CrossRefGoogle Scholar
 15.Xu Q, Anderson A, Gore J, Ding Z: Unified Bundling and Registration of Brain White Matter Fibers. IEEE Transactions on Medical Imaging. 2009, 28 (9): 13991411.CrossRefPubMedPubMedCentralGoogle Scholar
 16.Voineskos A, Rajji T, Lobaugh N, et al: Agerelated decline in white matter tract integrity and cognitive performance: A DTI tractography and structural equation modeling study. Neurobiology of Aging. 2010, doi:10.1016/j.neurobiolaging.2010.02.009,Google Scholar
 17.Wassermann D, Bloy L, Kanterakis E, Verma R, Deriche R: Unsupervised white matter fiber clustering and tract probability map generation: Applications of a Gaussian process framework for white matter fibers. Neuroimage. 2010, 51 (1): 228241. 10.1016/j.neuroimage.2010.01.004.CrossRefPubMedPubMedCentralGoogle Scholar
 18.Li H, Xue Z, Guo L, Liu T, Hunter J, Wong S: A hybrid approach to automatic clustering of white matter fibers. Neuroimage. 2010, 49: 12491258. 10.1016/j.neuroimage.2009.08.017.CrossRefPubMedGoogle Scholar
 19.Eckstein I, Shattuck D, Stein J, et al: Active fibers: Matching deformable tract templates to diffusion tensor images. Neuroimage. 2009, 47: 8289.CrossRefGoogle Scholar
 20.Zhukov L, Museth K, Breen D, Whitaker R, Barr A: Level set modeling and segmentation of DTMRI brain data. J Electron Imag. 2003, 12: 125133. 10.1117/1.1527628.CrossRefGoogle Scholar
 21.Vemuri B, Chen Y, McGraw T, Wang Z, Mareci T: Fiber tract mapping from diffusion tensor MRI. Proc IEEE Workshop on Variational and Level Set Methods in Computer Vision. 2001, 8188.CrossRefGoogle Scholar
 22.Wang Z, Vemuri B: Tensor field segmentation using region based active contour model. ECCV LNCS. 2004, 3024: 304315.Google Scholar
 23.Rousson M, Lenglet C, Deriche R: Level set and region based surface propagation for diffusion tensor MRI segmentation. Computer vision and mathematical methods in medical and biomedical image analysis. 2004, 123134.CrossRefGoogle Scholar
 24.Feddern C, Weickert J, Burgeth B: Level set methods for tensor valued images. 2003, 6572. Geometric and Level Set Methods in Computer Vision: In Proceedings of the Second IEEE Workshop on VariationalGoogle Scholar
 25.Jonasson L, Bresson X, Hagmann P, Cuisenaire O, Meuli R, Thiran J: White matter fiber tract segmentation in DTMRI using geometric flows. Med Imag Analysis. 2005, 3: 223236.CrossRefGoogle Scholar
 26.Jonasson L, Hagmann P, Pollo C, Bresson X, Wilson C, Meuli R, Thiran J: A level set method for segmentation of the thalamus and its nuclei in DTMRI. Signal Processing. 2007, 87: 309321. 10.1016/j.sigpro.2005.12.017.CrossRefGoogle Scholar
 27.Lenglet C, Rousson M, Deriche R: A statistical framework for DTI segmentation. IEEE Trans Med Imag. 2006, 25: 675700.CrossRefGoogle Scholar
 28.Arsigny V, Fillard P, Pennec X, Ayache N: LogEuclidean metrics for fast and simple calculus on diffusion tensors. Magnetic Resonance in Medicine. 2006, 56: 411421. 10.1002/mrm.20965.CrossRefPubMedGoogle Scholar
 29.Weldeselassie Y, Hamarneh G: DTMRI segmentation using graph cuts. Citeseer:. 65121, K65121Google Scholar
 30.Awate S, Hui Z, Gee J: A fuzzy, nonparametric segmentation framework for DTI and MRI analysis: With applications to DTItract extraction. IEEE Transactions on Medical Imaging. 2007, 26: 15251536.CrossRefPubMedGoogle Scholar
 31.Hasan K, EwingCobbs L, Kramer L, Fletcher J, Narayana P: Diffusion tensor quantification of the macrostructure and microstructure of human midsagittal corpus callosum across the lifespan. NMR in Biomedicine. 2008, 21: 10941101. 10.1002/nbm.1286.CrossRefPubMedPubMedCentralGoogle Scholar
 32.Jonasson L, Bresson X, Thiran J, Wedeen V, Hagmann P: Representing diffusion MRI in 5D simplifies regularization and segmentation of white matter tracts. IEEE Trans On Med Imaging. 2007, 26: 15471554.CrossRefGoogle Scholar
 33.NazemZadeh M, DavoodiBojd E, SoltanianZadeh H: Level set fiber bundle segmentation using spherical harmonic coefficients. Computerized Medical Imaging and Graphics. 2010, 34: 192202. 10.1016/j.compmedimag.2009.09.003.CrossRefPubMedGoogle Scholar
 34.NazemZadeh M, DavoodiBojd E, SoltanianZadeh H: AtlasBased Fiber Bundle Segmentation Using Principal Diffusion Directions and Spherical Harmonic Coefficients. NeuroImage. 2011, 54: 246164. Supplement 1CrossRefGoogle Scholar
 35.Descoteaux M, Deriche R: High angular resolution diffusion MRI segmentation using regionbased statistical surface evolution. Journal of Mathematical Imaging and Vision. 2008, 33 (2): 239252.CrossRefGoogle Scholar
 36.Hagmann P, Jonasson L, Deffieux T, Meuli R, Thiran J, Wedeen V: Fibertract segmentation in position orientation space from high angular resolution diffusion MRI. Neuroimage. 2006, 32: 665675. 10.1016/j.neuroimage.2006.02.043.CrossRefPubMedGoogle Scholar
 37.McGraw T, Vemuri B, Yezierski R, Mareci T: Segmentation of High Angular Resolution Diffusion MRI modeled as a field of von misesfisher mixtures. Proc European Conf on Computer Vision. 2006, 463475.Google Scholar
 38.Hess C, Mukherjee P, Han E, Xu D, Vigneron D: Qball reconstruction of multimodal fiber orientations using the spherical harmonic basis. Magn Reson Med. 2006, 56: 104117. 10.1002/mrm.20931.CrossRefPubMedGoogle Scholar
 39.Witelson S: Hand and sex differences in the isthmus and genu of the human corpus callosum: a postmortem morphological study. Brain. 1989, 112: 79910.1093/brain/112.3.799.CrossRefPubMedGoogle Scholar
 40.Lee S, Cheng J, Chen C, Tseng W: An automatic segmentation approach for boundary delineation of corpus callosum based on cell competition. Proc Engineering in Medicine and Biology Society Conf of the IEEE. 2008, 55145517.Google Scholar
 41.Osher S, Sethian J: Fronts propagating with curvature dependent speed: algorithms based on HamiltonJacobi formulations. J Comput Phys. 1988, 79: 1249. 10.1016/00219991(88)900022.CrossRefGoogle Scholar
 42.Osher S, Paragios N: Chapter: Level set methods. Geometric level set methods in imaging, vision, and graphics. SpringerVerlag New York Inc. 2003, 320.Google Scholar
 43.Dice LR: Measures of the amount of ecologic association between species. Ecology. 1945, 26: 297302. 10.2307/1932409.CrossRefGoogle Scholar
 44.Wakana S, Caprihan A, Panzenboeck M, et al: Reproducibility of quantitative tractography methods applied to cerebral white matter. Neuroimage. 2007, 36: 630644. 10.1016/j.neuroimage.2007.02.049.CrossRefPubMedPubMedCentralGoogle Scholar
 45.Zacharaki E, Hogea C, Shen D, Biros G, Davatzikos C: Nondiffeomorphic registration of brain tumor images by simulating tissue loss and tumor growth. Neuroimage. 2009, 46: 762774. 10.1016/j.neuroimage.2009.01.051.CrossRefPubMedPubMedCentralGoogle Scholar
Prepublication history
 The prepublication history for this paper can be accessed here:http://www.biomedcentral.com/14712342/12/10/prepub
Copyright information
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.