You are on page 1of 4

2015 7th International Conference on Intelligent Human-Machine Systems and Cybernetics

A New Brain MRI Image Segmentation Strategy


Based on K-means Clustering and SVM

Jianwei Liu Lei Guo


School of Automation School of Automation
Northwestern Polytechnical University Northwestern Polytechnical University
Xi’an, China Xi’an, China
xatu07@163.com lguo@ nwpu.edu.cn

Abstract—For the problem of noise and no reference image segmentation. However, the SVM cannot obtain the category
during brain magnetic resonance imagery (MRI) image of the classification data automatically, and selects training
segmentation, this paper proposes a new strategy to segment samples according to the reference segmentation image. In
brain MRI image based on K-means clustering algorithm and practice, generally there is no segmentation reference image
support vector machine (SVM).Firstly, the strategy segments when brain MRI image is segmented, therefore, selection of
brain MRI image using K-means clustering algorithm to training samples are a problem to be taken into account. In
obtain the initial classification result as the class label, this paper, in order to implement the unsupervised
secondly, the feature vectors of each pixel of brain tissue are classification, all brain image pixels are as the training
selected as the training samples and test samples, finally, brain
samples, which avoids artificial selection of training
MRI image is segmented by SVM. Experimental results show
samples, and the class label is obtained by K-means
that the proposed segmentation strategy obtains better
segmentation effect, especially has a good noise suppression for clustering algorithm.
brain images with low signal-noise-ratio (SNR). K-means clustering algorithm has been widely used in
medical image processing field because of the characteristics
Keywords-k-means clustering; support vector machine of simple, fast, unsupervised and good segmentation effect.
(SVM); feature extraction; The K-means clustering algorithm can divide brain MRI
image into white matter (WM), gray matter (WM) and
I. INTRODUCTION cerebrospinal fluid (CSF) and background automatically.
Although K-means clustering algorithm yields good results
Image segmentation plays an important role in medical for segmenting noise free images without require any manual
image processing, and is a key technique to analysis, intervention, but K-means clustering algorithm is more
understand and describe medical images in order to diagnose sensitive to noise, and it fails to segment images corrupted
curious diseases. At present brain magnetic resonance by noise and other imaging artifact[8][9].
imaging (MRI) segmentation has important theoretical and Therefore this paper proposes a new strategy to segment
application value in medical image analysis. Accurate brain brain MRI image based on K-means clustering algorithm and
tissue segmentation can improve reliability of brain disease support vector machine (SVM). Firstly, this paper uses K-
diagnosis and effectiveness of treatment. But MRI image is means algorithm to segment brain MRI image to get all
often affected by the noise, gray overlap, partial volume pixels category as the initial class label. Secondly, Good
effect, and the characteristic of different anatomical features which directly affect the segmentation effect are
structure, the classification of MRI image is a nonlinear selected for the SVM classification. Brain MRI image
classification problem[1][2][3]. At the same time the number contains the rich feature information, such as texture, gray
of brain images samples is limited, therefore, the currently etc. This paper adopts several important feature indexes as
popular classification method is support vector machine the input vector of SVM, such as gray, mean, variance,
(SVM) that V. Vapnik and C. Cortes proposed [4]. SVM is a entropy, gradient and mode. These features includes gray
kind of novel machine learning method, showing better information and spatial structure information of image
performance than the Bayesian classifier and neural network pixels, which help to reduce noise. Finally, brain image
classifier methods, and is regarded as a good alternative of segmentation image is obtained by using SVM training
the traditional learning classification method. Especially model and predicting brain image pixel category. The
SVM has good generalization performance in the case of simulation brain MRI image is used for brain tissue
small training samples and high dimension. Because of its classification, the experimental results show that the
excellent learning performance, SVM has become a hotspot proposed segmentation strategy obtains better segmentation
of machine learning field, and has good application prospect effect, especially has a good noise suppression for brain
and benefits to use it for MRI brain image processing images with low signal-noise-ratio (SNR).
research[5][6][7]. The rest of this paper is organized as follows. In the next
SVM is a supervised classifier, and needs to section segmentation strategy is introduced. Section 3
segmentation reference image during brain image

978-1-4799-8646-0/15 $31.00 © 2015 IEEE 270


DOI 10.1109/IHMSC.2015.182
presents experimental results and discussion. Finally, the This paper adopts support vector machine (SVM)
paper is summarized in section 4. package - LIBSVM, which was developed by Dr. Chih-
JenLin from Taiwan University, and has characteristics of
II. SEGMENTATION STRATEGY simple, fast, effective and general [10].
A. K-means Clustering Algorithm C. Feature Selection
K-means algorithm is a classical algorithm to solve the Segmentation result depends on features extraction to a
clustering problem proposed by MacQueen in 1967, and is large extent, in addition to the choice of the classifier. At
also a simple and efficient clustering algorithm. It can make present the common features include local gray statistical
the minimum square distance between all point in clustering features and local texture features, such as mean, median,
domain and the cluster center [8][9]. contrast, correlation, entropy, angular second moment
K-means algorithm classifies a sample set X (x1, x2, ···, xn) (ASM), and variance, etc. Let f denotes brain MRI image. i
into k clusters with the aim at minimizing an objective and j represents row and column of the image respectively.
function, where n is the number of samples and Ci represents Let f ' represents neighborhood image of the current pixel,
1 whose size is s x s, s is odd [6]. This paper adopts several
the clustering center, C = ¦ x , i =1, 2, ···, k, Ni is the
i main features as follows:
N i x∈xi
(1) Gray (I)
sample number of the ith cluster xi. The objective function J
is given as: I = f (i, j ) (2)
k n 2 (2) Mean ()
J = ¦¦ x (ji ) − Ci (1) 1 s−1 s−1 ಫ
i =1 j =1 μ= ¦¦ f (i, j )
s 2 i=0 j =0
(3)
2
(i )
Where, x j − Ci is a distance measure between a data (3) Standard deviation (Std)
s −1 s −1
1
point xj and the cluster center Ci, the Euclidean distance is
adopted as the distance measure in this paper. Therefore, the
Std =
s2
¦¦ ( f ಬ (i, j ) − μ )
i = 0 j =0
2
(4)
K-means algorithm is an iterative clustering algorithm that (4) Entropy (E)
finds a correct classification result which minimizes the sum L −1
squared error. The K-means algorithm is described as
follows:
E = −¦ pi log 2 pi (5)
i =0
(1) Initialize cluster centroids Ci with k random samples;
Where, E is the neighborhood image local entropy, pi is
(2) Assign each sample point xj to the nearest cluster
the ratio between the number of pixels, whose gray values
center.
are equal, and the total number of pixels in the image.
1 (5) Gradient (G)
(3) Recalculating each clustering center: C = ¦x, i
i
N i x∈xi
=1, 2, ···, k ; G (i, j ) = ( f (i + 1, j ) − f (i , j )) 2 + ( f (i , j + 1) − f (i , j )) 2 (6)
(4) Repeat steps (2) and (3) until Ci no longer changes.
B. SVM (6) Mode (M)
Mode represents the most frequency pixel gray value of
Brain MRI image classification is a small sample the neighborhood image.
learning problems, but the small sample learning exists lots
of problems such as over learning, lack of learning, which M (i, j ) = max{ frequency{ f ' (i, j )}} (7)
increases the difficulty of brain image classification. Support Then the above features are as the input vector of SVM.
vector machine (SVM) method adopts the structural risk Because the extracted features can have a lot of differences
minimization instead of the traditional empirical risk in values, especially ignoring small value features which
minimization, which not only obtains good classification contributed to the classification, the features normalization is
ability in the small sample, nonlinear and high dimensional very important. This paper uses the following formula to
space, but also has better extension. SVM is based on normalize features:
statistical learning theory of VC dimension (Vapnik - V
Chervonenk dimension) and structural risk minimization V= (8)
max(V )
principle. According to the limited sample information SVM
finds the best compromise between the complexity of the Here, V represents feature vector, max (V) represents
model and the learning ability, in order to get the best maximum value of V. After normalization, the feature values
generalization capability classification method [7]. are limited in the range of [0, 1].
Therefore, this paper uses SVM method to segment brain D. Brain MRI Image Segmentation Step Based on K-means
MRI images. and SVM
(1) Read brain MRI images for segmentation.

271
(2) Using K-means algorithm to segment brain MRI
image into four categories (WM, GM, CSF, background) as
the initial class label.
(3) Extracting the SVM training samples. In this paper, in
order to implement the unsupervised classification, all brain
image pixels are as the training samples and test samples,
which avoids artificial selection of training samples.
(4) Extracting features of samples, features of samples
are: gray, mean, variance, entropy, gradient and mode. Each
sample point corresponding to six image features and one
class label constitutes each pixel feature vector as training
samples to train SVM after normalization processing. Using
the same way, each sample point corresponding to six image
Figure 1. Simulation brain MRI Figure 2. Reference segmentaion
features constitutes each pixel feature vector as test samples image image
after normalization processing.
(5) Training samples from (4) and gets a SVM classifier,
namely the classification model, here this paper uses
Gaussian radial basis kernel function.
(6) Classifying the test samples from (4) by using
classification model from (5) to obtain segmentation image.
III. EXPERIMENTAL RESULTS
In order to verify the validity of the proposed strategy,
simulated brain MRI images are segmented by the proposed
strategy and K-means algorithm respectively. In order to
quantify the evaluation segmentation effect, this paper adopts
the Jaccard similarity (JS) coefficient analyze and compare
the segmentation result quantitatively [11], which is defined Figure 3. Segmentation results Figure 4. Segmentation results
as follows: by K-means algorithm by this strategy

V k ∩ VTk
J k ( S , T ) = Sk (9) TABLE I. THE EVALUATION OF BRAIN TISSUE SEGMENTATION
VS ∪ VTk Noise level Tissue K-means algorithm The proposed strategy
Where, VTk is the ground truth of class k, and VSk is the WM 0.8996 0.9086
segmental results, k represents brain tissue classification. The
ratio is between 0 and 1, the greater the value is, the better 5% GM 0.8074 0.8126
segmentation results get. CSF 0.7653 0.7343
The experimental data comes from brain MRI data of WM 0.8624 0.8928
McGill Montreal University Neurology Institute
(http://www.bic.mni.mcgill.ca/brainweb). The website 7% GM 0.7491 0.7914
provides a lot of MRI simulated brain data with a variety of CSF 0.7228 0.7330
slice thicknesses, noise levels, and levels of intensity non- WM 0.8122 0.8702
uniformity. The simulated brain MRI image database is often
9% GM 0.6721 0.7619
used as the gold standard for image segmentation. The size
of experiment data is 181 x 217 x 181, and its spatial CSF 0.6671 0.7300
resolution is 1 mm isotropic. The experiment adopts T1-
weighted data with 40% intensity non-uniformity and noise IV. CONCLUSION
level 5%, 7%, 9%. Here, 40% intensity non-uniformity and
In this paper, a new strategy to segment brain MRI image
noise level 9% T1-weighted simulated brain MRI image data
based on K-means clustering algorithm and support vector
as an example shows the segmentation results. Fig.1 is the
machine (SVM) was proposed for noise brain image
input image data, Fig.2 is reference segmentation image,
segmentation without reference image. Firstly, the class label
Fig.3 is segmentation result by K-means algorithm, and Fig.4
was obtained using K-means clustering algorithm, secondly,
segmentation results by the proposed segmentation strategy.
the feature vectors of each pixel of brain tissue are selected
Table I gives evaluation of brain tissue segmentation
as the training samples and test samples, finally, brain MRI
results of WM, GM, and CSF in terms of Jaccard Similarity
image is segmented by SVM. Experimental results showed
Coefficients for the simulated brain MRI images with
the proposed strategy obtained better segmentation effect,
different noise. As can be seen from Table I, the proposed
especially for the low SNR brain MRI images.
strategy obtains better segmentation accuracy than K-means
algorithm even at low SNR brain images.

272
ACKNOWLEDGMENT [6] H.X. Xu, Y. Li, G.X Zhu, X. Zhang and J.W. Tian, “Segmentation of
Magnetic Resonance Brain Tissues Image Based on Support Vector
This work is fully supported by a grant from the National Machines,” Journal of Image and Graphics, vol. 10, Oct. 2005, pp.
Natural Science Foundation of China (Project 1275-1280, doi:10.3969/j.issn.1006-8961.2005.10.012. ( in Chinese )
No.61273362). [7] H. Liu, K. Wang, and D. Liu, “Segmentation of Magnetic Resonance
Brain Tissues Image Based on Support Vector Machines and Level
REFERENCES Set Method,” Proc. Pattern Recognition, 2009. CCPR 2009, IEEE
Press, Nov. 2009, pp. 1-5, doi:10.1109/CCPR.2009.5344036. ( in
[1] A. Norouzi, M. S. M. Rahim, A. Altameem, et al, “Medical image Chinese )
segmentation methods, algorithms, and applications,” IETE Technical
Review, vol. 31, Jun. 2014, pp. 199-213, [8] K. Somasundaram1 and T. Genish, “Binarization of MRI with
doi:10.1080/02564602.2014.906861. Intensity Inhomogeneity using K-Means Clustering for Segmenting
Hippocampus,” International Journal of Multimedia and its
[2] A. Ortiz, J.M. Górriz, J.Ramírez, D. Salas-Gonzálezb, and J.M. Applications, vol. 5, Feb. 2013, pp. 11-18, doi:
Llamas-Elvirac, “Two fully-unsupervised methods for MR brain 10.5121/ijma.2013.5102.
image segmentation using SOM-based strategies,” Applied Soft
Computing, vol. 13, May 2013, pp. 2668-2682, doi: [9] M.K. Date and M.S.P. Akarte, “Brain image segmentation algorithm
10.1016/j.asoc.2012.11.020. using K-means clustering,” International Journal of Computer
Science And Applications, vol. 6, Dec. 2001, pp. 2127-2130.
[3] M. A. Balafar, A. R. Ramli, M. I. Saripan and S. Mashohor, “Review
of brain MRI image segmentation methods,” Artificial Intelligence [10] C. C. Chang and C.J. Lin, “LIBSVM: A library for support vector
Review, vol. 33, Mar. 2010, pp. 261-274, doi:10.1007/s10462-010- machines,” ACM Transactions on Intelligent Systems and
9155-0. Technology (TIST), vol. 2, Apr. 2011, pp. 27:1-27:27, doi:
10.1145/1961189.1961199.
[4] C. Cortes and V. Vapnik, “Support-vector Networks,” Machine
learning, vol. 20, Sept. 1995, pp. 273-297, doi:10.1007/BF00994018. [11] S.F. Wang, Z.Y. Geng, J.W. Zhang, Y.J. Chen and J. Wang, “A
Fuzzy C-means Model Based on the Spatial Structural Information
[5] J. Xiao, and Y. Tong, “Research of Brain MRI image segmentation for Brain MRI Segmentation,” International Journal of Signal
algorithm based on FCM and SVM,” Proc. Control and Decision Processing, Image Processing and Pattern Recognition, vol. 7, Feb.
Conference (2014 CCDC), IEEE Press, May 2014, pp. 1712-1716, 2014, pp. 313-322, doi: 10.14257/ijsip.2014.7.1.29.
doi:10.1109/CCDC.2014.6852445.

273

You might also like