Segmentation of MR image using local and global region based geodesic model
© Li et al.; licensee BioMed Central. 2015
Received: 21 October 2014
Accepted: 12 January 2015
Published: 19 February 2015
Segmentation of the magnetic resonance (MR) images is fundamentally important in medical image analysis. Intensity inhomogeneity due to the unknown noise and weak boundary makes it a difficult problem.
The paper presents a novel level set geodesic model which integrates the local and the global intensity information in the signed pressure force (SPF) function to suppress the intensity inhomogeneity and implement the segmentation. First, a new local and global region based SPF function is proposed to extract the local and global image information in order to ensure a flexible initialization of the object contours. Second, the global SPF is adaptively balanced by the weight calculated by using the local image contrast. Third, two-phase level set formulation is extended to a multi-phase formulation to successfully segment brain MR images.
Experimental results on the synthetic images and MR images demonstrate that the proposed method is very robust and efficient. Compared with the related methods, our method is much more computationally efficient and much less sensitive to the initial contour. Furthermore, the validation on 18 T1-weighted brain MR images (International Brain Segmentation Repository) shows that our method can produce very promising results.
A novel segmentation model by incorporating the local and global information into the original GAC model is proposed. The proposed model is suitable for the segmentation of the inhomogeneous MR images and allows flexible initialization.
Magnetic resonance image can provide excellent spatial resolution and superb soft tissue contrast for anatomical and functional structures. Accurate segmentation of MR image is an essential step in medical image analysis. However, the intensity inhomogeneity due to the unknown noise and weak boundary makes the segmentation a challenge. Various segmentation algorithms have been proposed in the literature. Particularly, Active contour model (ACM) receives the widespread attention since it can provide promisingly smooth and closed contours to cover object boundaries with sub-pixel accuracy [1, 2]. The existing ACMs can be mainly categorized into two classes: edge-based models and region-based models.
One of the most popular edge-based models is geodesic active contour (GAC) model [3, 4], which utilizes image gradient to construct an edge stopping function (ESF) to keep the contour evolution within the object boundaries. The model has been successfully applied in the general images with strong object boundaries, but it may suffer from boundary leakage in the brain MR images which typically contain weak boundaries due to low contrast and partial volume effect. However, when the initial contour is far away from the desired object boundary, the GAC model will fail to find the target . And then Song proposed an edge-based ACM that is driven by the regularized gradient flux flows . The method is not only robust to noise, but also preserves the edge information, thereby achieves accurate segmentation results.
Region-based models have many advantages over the edge-based ones. For example, region-based models are less sensitive to noise and contour initialization since they utilize region information as a substitute for image gradient to constrain contour evolution. Moreover, they can successfully segment images with the weak boundaries or even without boundaries. The well-known region-based model, Chan-Vese (CV) model, assumes that image intensities are statistically homogeneous in each region, and therefore it fails to segment MR images with intensity inhomogeneity . Then Li et al. proposed a local binary fitting (LBF) model to overcome the intensity inhomogeneity. The LBF model can provide desirable segmentation results because it uses the mean of the local region information. However, it is sensitive to the initial contours and is easy to trap into a local minimum, which limits their practical applications [8, 9]. Zhang et al. proposed a maximum likelihood in transformed domain method to simultaneously segment images and correct bias field. The method demonstrated the superiority by taking the mean and variance in a local region into account .
For making full use of the advantage of the three methods mentioned above, some hybrid models combining the local and global intensity fitting energies were proposed to drive the evolution of initial contours. Lei He et al. blended GMM model, Hueckel model and CV model, and finished the segmentation of inhomogeneous image, but is prone to be sensitive to initialization and parameterization . Zhang et al. proposed an improved method of GAC model, named as GCV model, which utilizes the global intensity information to construct a signed pressure force (SPF) function to drive the contour evolution. This method also proposed a new level set function re-initialization method, i.e., selective binary and Gaussian filtering regularized Level Set, which is robust and simple to implement, but is hard to deal with the image having inhomogeneous gray intensity or weak boundary images . Wang et al. proposed a hybrid level set method which has a LBF term based on the local intensity fitting and CV term based on an auxiliary global intensity fitting. Due to combining the local and global intensity information, the proposed model can avoid trapping into a local minimum .
Motivated by Zhang et al. and Wang et al. method, the paper proposes a novel method based on the GAC model. Concretely, a new SPF function is defined as an adaptive combination of the local and the global fitting terms. Local term is the local part being responsible for attracting the contour toward the object boundaries, and the global one is the auxiliary global part incorporating the global image information to drive the motion of the contour far away from object boundaries. Then the balance between the local and global fitting terms is dynamically adjusted by the weight calculated on the local intensity contrast. The proposed SPF function is regularized by a binary level set function to avoid the traditional re-initialization of the level set function to a signed distance function . The proposed model is first presented as a two-phase level set formulation and then extended to a multi-phase formulation. After minimizing the energy function , the method is able to segment MR images. Experiments on synthetic and MR images show that the proposed method can not only overcome inhomogeneous gray intensity, but also deal with the images having complex background and weak boundary. Moreover, it reduces function’s dependence on initial contour.
The rest of the paper is organized as follows. We briefly review several classic models and their limitation in “Background”. The proposed method is introduced in “Methods”. We discussed our proposed method and compared our segmentation results with those of GCV method, LBF method and Li et al. method  in “Results and discussion”. Finally, some conclusive remarks are included in “conclusion”
The related methods
Given the image I(x), x∈Ω and Ω ⊂ R 2. Let the initial contour C(q): → R 2 be a parameterized planar cure in Ω. In this subsection, we will introduce GAC model, CV model, GCV model and LBF, respectively.
The GAC model
where ∇K σ *I denotes convolving image I with a Gaussian kernel, K σ , whose standard deviation is σ. σ is the scale parameter that controls the region-scalability from the locally small neighborhood to the whole image domain and is adaptively chosen in the images, similar to .
The GAC model utilizes image gradient to construct an edge stopping function (ESF) to stop the contour evolution on the object boundaries. For images of weak boundaries or the initial contour is far away from the desired object boundary, the GAC model will fail to find the target .
The CV model
The GCV model
The GCV model shares the advantages of the CV and GAC models. This model utilizes the statistical information inside and outside the contour to construct a region-based SPF function to substitute ESF. Under the effect of SPF function, the contour can shrink when it is outside the object or expand when inside the object. Moreover, the level set function of the GCV model is regularized by the selective binary and Gaussian filtering , and the computational complexity is decreased by comparing with the traditional level set methods. However, this method can’t deal with the inhomogeneous or fuzzy boundary images. For example, Figure 1(c) shows that the GCV model fails to segment the object correctly.
The LBF model
Due to the localization property of the kernel function, the contribution of the intensity I(y) to the LBF energy decreases to zero as the point y moves away from the center point x. Therefore, the LBF energy is dominated by the intensity I(y) of points y in the neighborhood of x. This localization property plays a key role in segmenting the images with intensity inhomogeneity. Figure 2(b) illustrates the interaction of the local interior and local exterior regions of the point x. The local neighborhood of x, K σ (y - x), is represented by the black circle. The circle is spilt by the green curve into local interior (red) and local exterior (green) regions. The small yellow and blue dots represent the point x along the contour and point y in the local region of point x, respectively. f 1(x) and f 2(x) are computed in the local interior and local exterior region of the point x to fit the image intensities near the point x.
The design of novel SPF function
where R defines the local window of size 5 × 5 centered at x. ω varies between 0 and 1 and reflects how rapidly the intensity changes in a local region. It is smaller in the smooth regions and bigger in the regions close to the boundary of objects. Therefore, the ω can adaptively adjust the global term in all regions.
where α is the balloon force, which controls the contour shrinking or expanding.
Extension to multi-phase level set model
where ϵ is nonnegative constant. if ϵ is too small, the values of δ ϵ (z) tend to be near zero to make its effective range small, so the energy functional has a tendency to fall into a local minimum. The object may fail to be extracted if the initial contour starts far from it. However, if ϵ is large, although δ ϵ (z) tends to obtain a global minimum, the finial contour location may not be accurate. In “Results and discussion”, we will give some examples to show this drawback. We set ϵ = 0.3 for good approximation of δ by δ 3.
The main procedures of the proposed algorithm are summarized as follows:
where ρ > 0 is a constant, Ω0 is a subset in the image domain Ω and ∂Ω0 is the boundary of Ω0.
for Check whether the evolution of the level set function has converged
Compute C 1, C 2, f 1(x) and f 2(x) using (4), (5), (9) and (10), respectively.
Evolve the level set function according to Eq. (23).
Let ϕ = 1 if ϕ > 0; otherwise, ϕ = - 1.
Regularize the level set function with a Gaussian filter, i.e. ϕ = ϕ ∗ K σ
Results and discussion
Description of the parameters used in the study
To initialize the level set function. ρ > 0 is a constant.
Scale parameter in Gaussian kernel (GCV: σ = 1; LBF: σ = 3; our model: σ = 5).
Inner weight of contour C (LBF: λ 1 = 1).
Outer weight of contour C (LBF: λ 2 = 1).
The parameter of H 3 and δ 3 (LBF: ϵ = 1; our model: ϵ = 0.3).
Time step (GCV: ∆t = 1; LBF: ∆t = 0.1; our model: ∆t = 1).
The weight of ω (our model: β = 1).
π = 3.14 is a constant.
Balloon force (GCV and our model: α determined according to images).
Our method is first compared with the GCV and LBF model in the synthetic hand and the real blood images whose appearance show severe intensity inhomogeneity. And then our method is also compared in the real MR images to evaluate the performance of our method.
Iterations and CPU time (in seconds) needed by our model and LBF model when segmenting the image with the size of 110 × 111 in Figure 6 , respectively
Multi-phase segmentation of brain MR images
The segmentation of the brain MR images into WM, GM, and CSF is an important task in medical image analysis. A major difficulty in segmentation of MR images is the intensity inhomogeneity due to the noise. In this subsection, we will show an application of our multi-phase model to segment brain MR images. We also compare our method with the method of Li et al.  on 18 T1-w images obtained from the Internet Brain Segmentation Repository (IBSR) .
The average DSC values of WM, GM and CSF by the two methods, respectively
0.86 ± 0.03
0.83 ± 0.02
0.68 ± 0.14
0.89 ± 0.01
0.87 ± 0.02
0.63 ± 0.14
This paper presented a novel segmentation model by incorporating the local and global information into the original GAC model. Particularly, a new local SPF function is used to capture the local intensity information, so the novel model is especially fit for the segmentation of the inhomogeneous images. The weight balancing the global term is adaptively adjusted according to the statistics of the local intensity information. In a word, the proposed model can not only allow flexible initialization but also estimate intensity inhomogeneity. Moreover, the proposed method has better efficiency since it reduces the expensive re-initialization of the traditional level set method. In the future, the proposed method will be evaluated in the more extensive experiments.
This research was supported by the grants from National Natural Science Foundation of Chine grants 60972102, 81271670, 81471758, 81470868, the National High Technology Research and Development Program (2012AA02A606, 2015BAK31B01). This study was also supported by project 12441901600 of Science and Technology Commission of Shanghai Municipality and project 13XD1424800 of 2013 Shanghai Outstanding Technology Leaders Plan.
- Kass M, Witkin A, Terzopoulos D: Snakes: active contour models. Int J Comput Vision 1988,1(4):321–31. 10.1007/BF00133570View ArticleGoogle Scholar
- Lankton S, Tannenbaum A: Localizing region-based active contours. IEEE Trans Image Process 2008,17(11):2029–39.MathSciNetView ArticleGoogle Scholar
- Caselles V, Kimmel R, Sapiro G: Geodesic active contours. Int J Comput Vision 1997,22(1):61–79. 10.1023/A:1007979827043View ArticleMATHGoogle Scholar
- Goldenberg R, Kimmel R, Rivlin E, Rudzsky M: Fast geodesic active contours. IEEE Trans Image Process 2001,10(10):1467–75. 10.1109/83.951533MathSciNetView ArticleGoogle Scholar
- Zhang K, Zhang L, Song H, Zhou W: Active contours with selective local or global segmentation: a new formulation and level set method. Image Vision Comput 2010,28(4):668–76. 10.1016/j.imavis.2009.10.009View ArticleGoogle Scholar
- Song H: Active contours driven by regularised gradient flux flows for image segmentation. Electron Lett 2014,50(14):992–4. 10.1049/el.2014.1710View ArticleGoogle Scholar
- Chan T, Vese L: Active contours without edges. IEEE Trans Image Process 2001,10(2):266–77. 10.1109/83.902291View ArticleMATHGoogle Scholar
- Li C, Kao C, Gore JC, Ding Z: Implicit active contours driven by local binary fitting energy. Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) 2007, 1–7.Google Scholar
- Li C, Kao C, Gore JC, Ding Z: Minimization of region-scalable fitting energy for image segmentation. I IEEE Trans Image Process 2008,17(10):1940–9.MathSciNetView ArticleGoogle Scholar
- Zhang K, Liu Q, Song H, Li X: A Variational Approach to Simultaneous Image Segmentation and Bias Correction. 2014.Google Scholar
- He L, Zheng S, Wang L: Integrating local distribution information with level set for boundary extraction. J Vis Commun Image Represent 2010,21(4):343–54. 10.1016/j.jvcir.2010.02.009View ArticleGoogle Scholar
- Wang L, Li C, Sun Q, Kao C: Active contours driven by local and global intensity fitting energy with application to brain MR image segmentation. Comput Med Imag Grap 2009,33(7):520–31. 10.1016/j.compmedimag.2009.04.010View ArticleGoogle Scholar
- Aubert G, Kornprobst P: Mathematical problems in image processing: partial differential equations and the calculus of variations. Springer; 2006.Google Scholar
- Li C, Huang R, Ding Z, Gatenby J, Metaxas DN, Gore JC: A level set method for image segmentation in the presence of intensity inhomogeneities with application to MRI. IEEE Trans Image Process 2011,20(7):2007–16.MathSciNetView ArticleGoogle Scholar
- Xu C, Yezzi A Jr, Prince JL: On the relationship between parametric and geometric active contours. Proceedings of the Thirty-Fourth Asilomar Conference on Signals, Systems and Computers, vol 1 2000, 483–489.Google Scholar
- Yu Y, Zhang C, Wei Y, Li X: Active contour method combining local fitting energy and global fitting energy dynamically. Proceedings of the International Conference on Medical Biometrics, vol LNCS 6165163–72.Google Scholar
- Vese L, Chan T: A multiphase level set framework for image segmentation using the Mumford and Shah model. Int J Comput Vision 2002,50(3):271–93. 10.1023/A:1020874308076View ArticleMATHGoogle Scholar
- Chung G, Dinov I, Toga A, LA V: MRI tissue segmentation using a variational multilayer approach, Department of Mathematics, UCLA, CAM Report,08–54. 2008.Google Scholar
- Internet Brain Segmentation Repository. http://www.nitrc.org/projects/ibsr/
- Shattuck DW, Sandor-Leahy SR, Schaper KA, Rottenberg DA, Leahy RM: Magnetic resonance image tissue classification using a partial volume model. NeuroImage 2001,13(5):856–76. 10.1006/nimg.2000.0730View ArticleGoogle Scholar
- Valverde S, Oliver A, Cabezas M, Roura E, Lladó X: Comparison of 10 brain tissue segmentation methods using revisited IBSR annotations. J Magn Reson Imaging 2015, 41:93–101. 10.1002/jmri.24517View ArticleGoogle Scholar
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/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.