(IJACSA) International Journal of Advanced Computer Science and Applications,

Vol. 6, No. 7, 2015

An Improved Brain Mr Image Segmentation using

Truncated Skew Gaussian Mixture
Nagesh Vadaparthi
Srinivas Yerramalle Suresh Varma Penumatsa
Department of Information
Department of Information Department of Computer Science &
Technology
Technology Engineering
MVGR College of Engineering
GIT, GITAM University Adikavi Nannayya University
Vizianagaram, India
Visakhapatnam, India Rajahmundry, India

Abstract—A novel approach for segmenting the MRI brain Thus, in this paper Fuzzy C-Means (FCM) clustering
image based on Finite Truncated Skew Gaussian Mixture Model algorithm is considered for segmenting the image into number
using Fuzzy C-Means algorithm is proposed. The methodology is of regions and derive the model parameters. The obtained
presented evaluated on bench mark images. The obtained results parameters are thus refined further using the EM algorithm.
are compared with various other techniques and the
performance evaluation is performed using Image quality The rest of the paper is organized as follows: section-2
metrics and Segmentation metrics. explains about the FCM algorithm, section-3 deals with the
concept of Finite Truncated Skew Gaussian distribution and
Keywords—Truncated Skew Gaussian Mixture model; section-4 handles the initialization of parameters. Section-5
Segmentation; Image quality metrics; Segmentation metrics; Fuzzy shows the updating of parameters and section-6 demonstrates
C-Means clustering the proposed segmentation algorithm. In section-7 the
experimental results are discussed, section-8 concludes the
I. INTRODUCTION paper, the scope for further enhancement is proposed in section
MRI segmentation plays a vital role in medical research 9 of the paper.
and applications. MRI has wide range of advantages over other
conventional imaging techniques since magnetization and radio II. FUZZY C-MEANS CLUSTERING ALGORITHM
waves are used instead of X-rays in making the detailed and The first step in any segmentation algorithm is to divide
cross-sectional images of the brain [1]. Various operations image into different image regions. Many segmentation
based on image processing were defined earlier on MR images. algorithms are presented in literature [10],[11],[12],[13],[14].
Among these, segmentation of brain images into sub-regions Among these techniques, medical image segmentation based
has enormous research and medical applications. These sub- on K-Means is mostly utilized [4]. But, the main disadvantage
regions are utilized in visualizing and analyzing the anatomical with K-Means is that, K-Means are slow in convergence and
structures in the brain which help in neuro-surgical planning pseudo unsupervised learning that requires the initial value of
[2]. K. Apart from K-Means, hierarchical clustering algorithm[5] is
There are various conventional methods for MRI also used but even this algorithm shares similar arguments as
segmentation which require human interaction in terms of the case of K-Means algorithm. Fuzzy C-Means clustering
mentioning number of classes for obtaining accurate and algorithm in considered, in order to identify the initial clusters.
reliable segmentation. Thus, it is necessary to derive new The algorithm for Fuzzy C-means clustering is presented
techniques for effective segmentation. Much of the emphasis below.
has been given to the segmentation algorithm based on finite The FCM employs fuzzy partitioning such that a data point
normal mixture models where each image is assumed to be a can belong to all groups with different membership grades
mixture of Gaussian distributions. But actually it is observed between 0 and 1 and it is an iterative algorithm. The aim of
that the pixels are quantized through the brightness or contrast FCM is to find cluster centers (centroids) that minimize a
in the gray scale level (Z) at that point. It is also observed that dissimilarity function. To accommodate the introduction of
the image regions have finite range of pixel intensities (-∞, +∞) fuzzy partitioning, the membership matrix (U) is randomly

u
and may not be symmetric and Meso kurtic [3]. In this paper, initialized according to Equation (1).

 1, j  1,..., n
to have an accurate modeling of the feature vector, finite c
truncated skew Gaussian is considered by assuming that the
(1)
i 1
pixel intensities in the entire image follow a Finite Truncated ij
Skew Gaussian distribution [4][5][6][7][8].
Hence, in order to segment more accurately Fuzzy C- The dissimilarity function which is used in FCM is given
Equation (2)

J (U , c1 , c2 ,..., cc )   J i   uij dij

Means algorithm is preferred because of the additional
flexibility that allows the pixel to belong to multiple classes c c n
with varying degree of membership [9]. m 2
(2)
i 1 i 1 j 1

190 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 6, No. 7, 2015

Where, uij is between 0 and 1; and with this very assumption, the classification process is
carried out. But in reality, white matter regions contain certain
ci is the centroid of cluster i; portion of grey matter at the boundaries and the tissues within
dij is the Euclidian distance between ith centroid(ci) and jth these regions are assumed to contain the pixels having the
data point; probabilities which may be both symmetric and non-symmetric
[5]. The problem gets multifold in case of abnormal brains,
m є [1,∞] is a weighting exponent. since registering these images with prior probabilities is
To reach a minimum of dissimilarity function there are two difficult, as each pixel inside a region may belongs to a

 j 1 uij x j
conditions. These are given in Equation (3) and Equation (4). different class.
The effect of partial volume is due to the assumption of

ci 
n


m
considering the distribution of the pixels inside the image
n
(3) regions as normal. Hence it is necessary to consider
m
j 1 ij
u asymmetric distributions as the brain can diverge from the
symmetric population, in order for it to be segmented

uij 
1 satisfactorily.

d 
k 1  d ij
2 /( m 1)
(4)

The crucial information regarding the deformities in the


c brain can be available from the segmented regions, which may
 kj 
be skewed. In most of the brain related data, the information
about the damaged tissues may be located at the boundaries
This algorithm determines the following steps. (outliers) and the pixels inside these regions may exhibit non
homogenous features, which include asymmetry,
Step-1:Randomly initialize the membership matrix (U) that multimodality exhibiting long tails.
has constraints in Equation (1).
Hence to have an effective analysis about the damaged
Step-2: Calculate centroids (ci) by using Equation (3). tissues, one need to consider mixture models which can
Step-3:Compute dissimilarity between centroids and data accommodate data having non-normal features. Notable
points using equation (2). Stop if itsimprovement over previous distribution among such models include the skew normal
iteration is below a threshold. mixture model [15][16], the skew t -mixture model
[17][18][19], the skew t –normal mixture model [20], and some
Step-4: Compute a new U using Equation (4). Go to Step 2. other non-elliptical approaches [21][22][23]. The log-Normal,
the Burr, the Weibull the Gamma and the Generalized Pareto
By iteratively updating the cluster centers and the
distribution are also considered in the literature for analyzing
membership grades for each data point, FCM iteratively moves
asymmetric data. Among these models, to model the data
the cluster centers to the "right" location within a data set.
having long tails efficiently Skew Gaussian Mixture models
FCM does not ensure that it converges to an optimal solution.
are more appropriate [24][25][26].
Because of cluster centers (centroids) are initialized using U
that randomly initialized (Equation (3)). Skew symmetric distributions are mainly used for the set of
images where the shape of image regions are not symmetric or
Performance depends on initial centroids. For a robust
bell shaped distribution and these distributions can be well
approach there are two ways which is described below.
utilized for the medical images where the bone structure of the
1) Using an algorithm to determine all of the centroids humans are asymmetric in nature. To have a more accurate
(for example: arithmetic means of all data points). analysis of the medical images, it is customary to consider that
2) Run FCM several times each starting with different in any image, the range of the pixels is finite in nature. Among
initial centroids. the pixels generated from the brain images, to extract the
features effectively only finite ranges of pixels are very much
III. FINITE TRUNCATED SKEW GAUSSIAN DISTRIBUTION useful. Hence, to have a more closure and deeper
approximation of the medical data, truncated skew normal
In any medical image, pixel is used as a measure of distribution are well suited
quantification and the entire medical image is assumed as a
heterogeneous collection of pixels and each pixel is influenced The probability density function of the truncated skew
by various factors such as brightness, contrast, saturation etc. normal distribution is given by
For effectual analysis and classification of the brain tissues, it
is obligatory to uphold good contrast between white matter and (5)
grey matter, but in general the grey matter structure consists of
tissues with varying intensities compared to that of white where, µ ϵ R, σ > 0 and λ ϵ R represents the location, scale
matter, thereby making it a challenging task for effective and shape parameters respectively. Where and � denote the
classification of tissues in these regions. probability density function and the cumulative density
function of the standard normal distribution.
A further issue associated is the problem of partial volume,
many models assume that the pixels inside a particular tissue The limits and of the truncated normal distribution are Zl
have homogenous properties, and follow a symmetric pattern =a and Zm = b. Where Zl and Zm denotes the truncation limits.

191 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 6, No. 7, 2015

Truncating equation (1) between these limits, the following VI. SEGMENTATION ALGORITHM
equations are obtained After refining the estimates, the important step is to convert
the heterogeneous data into homogenous data or group the
∫ (6)
related pixels. This process is carried out by performing the
where, segmentation. The image segmentation is done in 3 steps:
∫ (7) Step-1: Obtain the initial estimates of the finite truncated
and skew Gaussian mixture model using Fuzzy C-Means
Clustering algorithm.
∫ (8) Step-2: Using the initial estimates obtained from step-1, the
where, EM algorithm is iteratively carried out.
fµ, σ, λ (x) is as given in equation (1) Step-3: The image segmentation is carried out by assigning
each pixel into a proper region (Segment) according to
∫ ̅ ∫ ̅ (9) maximum likelihood estimates of the jth element Lj according
to the following equation
IV. INITIALIZATION OF PARAMETERS
In order to initialize the parameters, it is needed to obtain {∫ ̅ ∫ ̅ }
the initial values of the model distribution. The estimates of the
Mixture model µi, σi and αi where i=1, 2.., k are estimated VII. EXPERIMENTAL RESULTS & PERFORMANCE
using Fuzzy C-Means Clustering algorithm as proposed in EVALUATION
section-II. It is assumed that the pixel intensities of the entire
In order to evaluate the performance of the developed
image is segmented into a K component model πi, i=1, 2...K
algorithm, T1 weighted images were used. The input medical
with the assumption that πi = 1/K where K is the value
images are obtained from brain web images. It is assumed that
obtained from Fuzzy C-Means Clustering algorithm discussed
the intensities of the pixels in medical images are asymmetric
in section-2.
in nature. Hence, follow a skew Gaussian distribution and as
V. UPDATING INITIAL ESTIMATES THROUGH EM the limits are finite and within the specified range of values are
ALGORITHM only necessary in medical image segmentation process
Truncated skew Gaussian distribution. Is used The
The initial estimates of µi, σi and αi that are obtained from initialization of parameters for each segment is achieved by
section – 4 are to be refined to obtain the final estimates. For using Fuzzy C-Means Clustering algorithm and the estimates
this purpose EM algorithm is utilized. The EM algorithm are updated using the EM algorithm. The experimentation is
consists of 2 steps E-step and M-Step. In the E-Step, the initial carried out by using the segmentation algorithm depicted in
estimates obtained in section – 4 are taken as input and the section-6 and the obtained results are evaluated using
final updated equations are obtained in the M-Step. The segmentation quality metrics[27] such as Jacquard Coefficient
updated equations for the model parameters µ, σ and α are (JC), Volumetric Similarity (VS), Variation of Information
given below. (VOI), Probabilistic Rand Index (PRI) and Global Consistency
∫ ̅ ∫ ̅
Error (GCE) and the formulas for calculating these metrics are
(10) given as follows:
∫ ̅ ∫ ̅
| |
∫ ̅ ∫ ̅
| |
(13)
∫ ̅ ∫ ̅ || | | || | |
| | | |
(14)
(11)
∫ ̅
Where, a=|X ∩Y |, b= |X/Y|, c=|Y/X|, d |̅̅̅̅̅̅̅̅|
∫ ̅ ∑ ∑ (15)
∫ ̅ | |
Where, LRE = | |
S and S’ are segment classes
{ ∫ ̅ }
⁄
(12) and xi is the pixel.
∫ ̅

∫ ̅ VOI (X,Y)= H(X) = H (Y) – 2I(X;Y) (16)

∫ ̅ Where, X and Y are two clusters
[ ∫ ̅ ]
PRI(St,{S})=( ) ∑ [ ( ) ( ) ] (17)
∫ ̅

∫ ̅ Where, ( ) ∑ and the

∫ ̅ values range from 0 to 1. The value 1 denotes the segments are
[ ∫ ̅ ] identical.

192 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 6, No. 7, 2015

TABLE I. SEGMENTATION QUALITY METRICS

Skew
Truncated Skew Skew
GMM Truncated Truncated
Quality SGMM GMM GMM Standard Standard
Image GMM with K- SGMM with SGMM
Metric with K- with with FCM Limits Criteria
Means- HC with FCM
Means HC-EM -EM
EM
Close to 1
JC 0.089 0.689 0.711 0.703 0.736 0.795 0.832 0 to 1
Close to 1
VS 0.432 0.733 0.781 0.8799 0.887 0.891 0.923 0 to 1
Possible
B0S1 VOI 2.3665 5.3173 5.2323 5.142 5.381 5.232 4.7099 -∞ to ∞
Big
GCE 0.2802 0.5964 0.6088 0.561 0.626 0.4223 0.5025 0 to 1
Close to 1
PRI 0.504 0.6396 0.6697 0.619 0.663 0.7958 0.6009 0 to 1
Close to 1
Close to 1
JC 0.0677 0.7656 0.7921 0.7921 0.812 0.819 0.851 0 to 1
Close to 1
VS 0.3212 0.8767 0.8801 0.8814 0.892 0.8914 0.923 0 to 1
Possible
B0S2 VOI 1.9724 3.924 0 4.35 4.63 6.2894 4.9823 -∞ to ∞
Big
GCE 0.2443 0.4741 0 0.419 0.5013 0.4664 0.5125 0 to 1
Close to 1
PRI 0.416 0.5016 1 0.514 0.542 0.6847 0.6506 0 to 1
Close to 1
Close to 1
JC 0.0434 0.6567 0.689 0.7143 0.722 0.784 0.818 0 to 1
Close to 1
VS 0.123 0.812 0.849 0.916 0.932 0.926 0.947 0 to 1
Possible
B0S3 VOI 0.7684 0.2916 0 1.659 2.956 5.5318 4.3623 -∞ to ∞
Big
GCE 0.089 0.031 0 0.107 0.02 0.4001 0.3943 0 to 1
Close to 1
PRI 0.576 0.5853 1 0.632 0.661 0.706 0.7111 0 to 1
Close to 1
Close to 1
JC 0.0456 0.7878 0.7891 0.874 0.896 0.911 0.933 0 to 1
Close to 1
VS 0.2233 0.3232 0.465 0.54 0.621 0.643 0.722 0 to 1
Possible
B0S4 VOI 1.268 1.569 0 3.354 3.693 4.1619 2.9053 -∞ to ∞
Big
GCE 0.056 0.091 0 0.157 0.199 0.2949 0.2554 0 to 1
Close to 1
PRI 0.189 0.191 1 0.496 0.519 0.5628 0.6987 0 to 1
Close to 1
Close to 1
JC 0.141 0.776 0.779 0.791 0.8123 0.826 0.861 0 to 1
Close to 1
VS 0.313 0.397 0.452 0.784 0.797 0.7910 0.811 0 to 1
Possible
B1S1 VOI 1.6499 4.0874 3.9136 3.951 4.13 4.4115 3.5797 -∞ to ∞
Big
GCE 0.1874 0.4487 0.4651 0.418 0.4468 0.2752 0.4103 0 to 1
Close to 1
PRI 0.9256 0.6678 0.7578 0.6258 0.6692 0.686 0.8044 0 to 1
Close to 1
Close to 1
JC 0.098 0.7892 0.7902 0.877 0.908 0.896 0.912 0 to 1
Close to 1
VS 0.0433 0.878 0.898 0.881 0.896 0.918 0.931 0 to 1
Possible
B1S2 VOI 2.3215 2.8047 2.921 3.91 5.122 6.6411 2.8047 -∞ to ∞
Big
GCE 0.2838 0.3407 0.348 0.339 0.3695 0.4661 0.3407 0 to 1
Close to 1
PRI 0.3807 0.369 0.429 0.485 0.561 0.6322 0.8690 0 to 1
Close to 1
Close to 1
JC 0.0222 0.8926 0.899 0.9124 0.9236 0.946 0.969 0 to 1
Close to 1
VS 0.3223 0.3429 0.425 0.3543 0.359 0.3869 0.441 0 to 1
Possible
B1S3 VOI 1.2411 0.9988 1.252 2.665 3.6351 6.7129 0.9988 -∞ to ∞
Big
GCE 0.1466 0.1157 0.227 0.398 0.424 0.4559 0.1157 0 to 1
Close to 1
PRI 0.9576 0.9662 0.856 0.652 0.698 0.7202 0.9675 0 to 1
Close to 1
0.455 Close to 1
JC 0.762 0.797 0.815 0.826 0.854 0.889 0 to 1
0.329 Close to 1
VS 0.7001 0.779 0.7158 0.754 0.786 0.895 0 to 1
-8.8e- Possible
B1S4 VOI 0.201 1.332 0.19 2.35 5.0898 5.561 -∞ to ∞
16 Big
GCE 0.112 0.176 0.212 0.265 0.3062 0.5214 0 to 1
0.119 Close to 1
PRI 0.1001 0.129 0.27 0.353 0.5573 0.691 0 to 1
0.065 Close to 1

193 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 6, No. 7, 2015

1.2 GMM 0.7 GMM

1 0.6
Skew GMM 0.5 Skew GMM
0.8
with K-Means- 0.4 with K-Means-
0.6 EM 0.3 EM
Truncated Truncated
0.4 0.2
SGMM with K- SGMM with K-
0.2 Means 0.1 Means
Skew GMM 0 Skew GMM
0 with HC-EM

B1S3
B0S1
B0S2
B0S3
B0S4
B1S1
B1S2

B1S4
with HC-EM
B0S1 B0S3 B1S1 B1S3
a) Jacquard Coefficient d) Global Consistency Error
1 1.2 GMM
GMM
0.8 1
Skew GMM 0.8 Skew GMM
0.6 with K-Means-
with K-Means- 0.6
0.4 EM EM
Truncated 0.4 Truncated
0.2 SGMM with K- SGMM with K-
0.2
Means Means
Skew GMM 0 Skew GMM
0
with HC-EM

B0S1
B0S2
B0S3
B0S4
B1S1
B1S2
B1S3
B1S4
B0S1 B0S3 B1S1 B1S3 with HC-EM

b) Volume Similarity
e) Probabilistic Rand Index
8 GMM Fig. 1. Segmentation Quality Metrics
7
6 The reconstruction process is carried out by positioning
5 Skew GMM each pixel into its appropriate location. The performance
with K-Means- evaluation [27] of the obtained output is done using the image
4
EM quality metrics such as Average difference, Maximum
3 Truncated distance, Image Fidelity, Means Squared Error and Peak
2 SGMM with K- Signal-to-Noise ratio.
1 Means
Skew GMM The developed algorithm is compared with Skew Gaussian
0 mixture model with K-Means, Hierarchical Clustering, Fuzzy
with HC-EM
B0S1
B0S2
B0S3
B0S4
B1S1
B1S2
B1S3
B1S4

-1 C-Means, Truncated Skew Gaussian mixture model with K-

Means and Hierarchical Clustering algorithms and the results
c) Variation of Information obtained are tabulated in Table-1, Table-2 and fig.-1 & fig-2.

TABLE II. SEGMENTATION QUALITY METRICS

Skew Truncated Skew
GMM SGMM GMM Truncated SGMM Truncated
Quality Standard Standard
Image GMM with with K- with HC SGMM with with Fuzzy SGMM
Metric Limits Criteria
K- Means HC CMean with FCM
Means
AD 0.573 0.773 0.792 0.812 0.835 0.8451 0.899 -1 to 1 Closer to 1
MD 0.422 0.922 0.941 0.9325 0.939 0.945 0.973 -1 to 1 Closer to 1
B0S1 IF 0.416 0.875 0.428 0.923 0.941 0.9756 0.9805 0 to 1 Closer to 1
MSE 0.04 0.134 2.19e-005 0.094 2.92E-005 9.3E-07 3.03e-005 0 to 1 Closer to 0
SNR 17.41 29.23 72.15 33.89 87.39 108.42 93.324 -∞ to ∞ Possible Big

194 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 6, No. 7, 2015

AD 0.37 0.876 0.887 0.749 0.798 0.49 0.921 -1 to 1 Closer to 1

MD 0. 221 0.897 0.910 0.912 0.926 0.931 0.951 -1 to 1 Closer to 1
B0S2 IF 0.336 0.876 0.894 0.859 0.873 0.9046 0.991 0 to 1 Closer to 1
MSE 0.240 0.211 0.124 0.2019 0.102 3.6E-06 3.01e-005 0 to 1 Closer to 0
SNR 14.45 35.65 84.23 39.85 89.65 102.5 93.34 -∞ to ∞ Possible Big

AD 0.456 0.76 0.796 0.81 0.826 0.6721 0.835 -1 to 1 Closer to 1

MD 0.345 0.879 0.847 0.807 0.86 0.911 0.947 -1 to 1 Closer to 1
B0S3 IF 0.44 0.86 0.883 0.917 0.919 0.9366 0.928 0 to 1 Closer to 1
MSE 0.22 0.23 0.2012 0.2123 0.267 2.43E-06 3.55e-005 0 to 1 Closer to 0
SNR 19.88 37.98 77.46 39.71 82.31 104.27 92.63 -∞ to ∞ Possible Big

AD 0.231 0.473 0.5023 0.4991 0.612 0.7731 0.79 -1 to 1 Closer to 1

MD 0. 224 0.977 0.954 0.971 0.977 0.9001 0.996 -1 to 1 Closer to 1
B0S4 IF 0.212 0.813 0.889 0.892 0.882 0.8835 0.929 0 to 1 Closer to 1
MSE 0.24 0.121 0.1012 0.1192 1.02E-05 4.46E-06 2.72e-005 0 to 1 Closer to 0
SNR 21.42 33.28 35.6 37.41 78.8 101.634 93.79 -∞ to ∞ Possible Big

AD 0.342 0.764 0.7661 0.7015 0.794 0.6957 0.861 -1 to 1 Closer to 1

MD 0.317 0.819 0.919 0.854 0.921 0.815 935 -1 to 1 Closer to 1
B1S1 IF 0.391 0.812 0.856 0.876 0.898 0.985 0.991 0 to 1 Closer to 1
MSE 0.251 0.228 1.34e-005 0.1759 2.64E-005 4.62E-07 7.87e-006 0 to 1 Closer to 0
SNR 3.241 5.514 32.154 5.68 89.31 111.482 99.173 -∞ to ∞ Possible Big

AD 0.21 0.3653 0.654 0.232 0.661 0.4596 0.712 -1 to 1 Closer to 1

MD 0.21 0.892 0.8825 0.912 0.921 0.891 0.913 -1 to 1 Closer to 1
B1S2 IF 0.213 0.787 0.813 0.791 0.851 0.7893 0.958 0 to 1 Closer to 1
MSE 0.06 0.145 0.096 0.594 0.024 6.49E-06 1.31e-006 0 to 1 Closer to 0
SNR 13.43 49.22 99 20.39 99 100.001 106.95 -∞ to ∞ Possible Big

AD 0.323 0.322 0.554 0.4592 0.54 0.4398 0.59 -1 to 1 Closer to 1

MD 0.123 0.212 0.413 0.456 0.446 0.546 0.596 -1 to 1 Closer to 1
B1S3 IF 0.233 0.897 0.917 0.923 0.926 0.915 0.99 0 to 1 Closer to 1
MSE 0.01 0.4345 0.002 0.119 1.29E-005 2.62E-06 2.16e-007 0 to 1 Closer to 0
SNR 11.11 27.267 39.12 29.86 71.69 103.95 114.78 -∞ to ∞ Possible Big

AD 0.314 0.338 0.635 0.497 0.699 0.521 0.732 -1 to 1 Closer to 1

MD 0.241 0.249 0.294 0.317 0.391 0.452 0.569 -1 to 1 Closer to 1
B1S4 IF 0.293 0.683 0.697 0.791 0.781 0.8756 1 0 to 1 Closer to 1
MSE 0.18 0.197 0.113 0.213 0.829 3.83E-06 0.023 0 to 1 Closer to 0
SNR 21.21 78.19 99 99 99 102.2932 106.26 -∞ to ∞ Possible Big

GMM
1 1.2 GMM
Skew GMM with K- 1
0.8 Means Skew GMM with
Truncated SGMM K-Means
0.8 Truncated SGMM
0.6 with K-Means
Skew GMM with with K-Means
0.6 Skew GMM with
0.4 HC
Truncated SGMM HC
0.4 Truncated SGMM
0.2 with HC with HC
SGMM with Fuzzy 0.2 SGMM with Fuzzy
0 CMean CMean
Truncated SGMM 0
B1S3
B0S1
B0S2
B0S3
B0S4
B1S1
B1S2

B1S4

Truncated SGMM
with FCM B0S1 B0S3 B1S1 B1S3 with FCM
a) Average Difference b) Maximum Distance

195 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 6, No. 7, 2015

The proposed method is compared to the methods based on

1.2
GMM Finite Gaussian Mixture Model, Finite Skew Gaussian Mixture
Model with K – Means algorithms, Finite Skew Gaussian
1 Mixture Model with Hierarchical Clustering algorithms and
Skew GMM with K-
Means Finite Truncated Skew Gaussian Mixture Model with K –
0.8 Means and Hierarchical Clustering algorithms. The
Truncated SGMM
segmentation algorithm so developed is applied to 8 sub-
with K-Means images as White Matter (WM), Gray Matter (GM), Cerebro
0.6
Skew GMM with Spinal Fluid (CSF) and Background of 2 brain images namely
HC B0S1, B0S2, B0S3, B0S4, B1S1, B1S2, B1S3 and B1S4. The
0.4 segmentation algorithm is developed and the performance of
Truncated SGMM
with HC the segmentation algorithm is evaluated through segmentation
0.2 quality metric such as jacquard Coefficient (JC), Volumetric
SGMM with Fuzzy
Similarity (VS), Variation of Information (VOI), Global
CMean
0 Consistency Error (GCE) and Probabilistic Random Index
Truncated SGMM (PRI). The values after segmentation by using above quality
B0S1
B0S2
B0S3
B0S4
B1S1
B1S2
B1S3
B1S4

with FCM metric are presented in Table-1. From the above table, for the
medical image B0S1, the values of the JC and VS, the values
c) Image Fidelity of the developed method are close to 1 which implies that the
0.9 segmentation methodology that is developed outperforming the
GMM segmentation model developed by using Gaussian Mixture
0.8 Model and Finite Skew Gaussian Mixture model using K-
Skew GMM with Means algorithm. The other metrics such as VOI, GCE and
0.7 PRI also are superior in the developed model when compared
K-Means
to the existing model for medical image segmentation based on
0.6 Truncated SGMM medical image segmentation using Finite Truncated Skew
0.5 with K-Means Gaussian Mixture Models.
Skew GMM with From the above tables-1 & 2 and the fig.-1 and fig.-2, it is
0.4 HC observed that the performance of medical image segmentation
0.3 Truncated SGMM based on Finite Truncated Skew Gaussian Mixture Model
with HC using Fuzzy C-Means algorithm, the Average Difference (AD)
0.2 for the image B0S1 is closure to 1 when compared to that of
SGMM with Fuzzy Gaussian Mixture Model, Skew Gaussian Mixture Model with
0.1 CMean K-Means, Hierarchical Clustering & Fuzzy C-Means and
Truncated SGMM Finite Truncated Skew Gaussian Mixture Model with K-Means
0 with FCM & Hierarchical Clustering algorithms. Similarly, the other
B0S1
B0S2
B0S3
B0S4
B1S1
B1S2
B1S3
B1S4

quality metrics such as Maximum Distance (MD), Image

Fidelity (IF), Mean Squared Error (MSE) and Signal to Noise
d) Mean Square Error Ratio (SNR) are more superior for the developed method than
that of the model based on Gaussian Mixture Model. This can
140 GMM be clearly seen from the output images given in Graph-6.3. The
same phenomenon is observed for the other medical images
120 B0S1, B0S2, B0S3, B0S4, B1S1, B1S2, B1S3 and B0S4.
Skew GMM with
100 K-Means In all these images there is a drastic improvement in Image
Truncated SGMM Quality metrics, the edges in medical image are more clearly
80 with K-Means visible. In the developed method, when compared to GMM,
Skew GMM with K-Means, Hierarchical Clustering & Fuzzy
Skew GMM with C-Means and Truncated Skew GMM with K-Means &
60 HC Hierarchical Clustering algorithms, the signal to noise ratio has
40 Truncated SGMM increased and the Average Difference & Image Fidelity are
with HC close to 1 and Mean Squared Error is close to 0 which implies
20 SGMM with Fuzzy that in the developed method, the edges are more closely
CMean visible and since MSE is much closure to 0, the output image is
0 more closure to input image. Thus, the developed algorithm
Truncated SGMM has the advantage that since the edges are much clearer, it gives
B0S1
B0S2
B0S3
B0S4
B1S1
B1S2
B1S3
B1S4

with FCM a very comprehensive idea regarding the details of the medical
images. The developed model helps to analyze the medical
e) Signal to Noise Ratio images in a better contrast than that of the existing models.
Fig. 2. Image Quality Metrics

196 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 6, No. 7, 2015

VIII. CONCLUSION [8] Nagesh Vadaparthi, Srinivas Yerramalle, and Suresh Varma.P: On
Improved Medical Brain MR Image Segmentation Based on Truncated
This proposed article is focused towards MRI Brain Image Skew Gaussian Mixture Model using Hierarchical Clustering and EM
Segmentation. A new approach based on Finite Truncated algorithms”, International Journal of Advanced Research in Computer
Skew Gaussian Mixture Model is introduced. The performance Science (ISSN: 0976-5697) , 2(6) November-December 2011.
evaluation of the developed model is investigated by using [9] M. Sashidhar et al: MRI Brain Image Segmentation using Modified
Fuzzy C-Means Clustering Algorithm, IEEE-Int. Conf. on
Image quality metrics which depict that the developed Communication systems and Network Technologies, 2011, Pp.473-478.
algorithm outperforms the other existing algorithms based on
[10] D. L. Pham, C. Y. Xu, and J. L. Prince: A survey of current methodsin
Skew Gaussian mixture model using K-Means, Skew Gaussian medical image segmentation, Annu. Rev. Biomed. Eng., vol. 2, pp.315–
mixture model Hierarchical Clustering, Skew Gaussian mixture 337, 2000.
model Fuzzy C-Means, Truncated Skew Gaussian mixture [11] K. Van Leemput, F. Maes, D. Vandeurmeulen, and P. Suetens:
model with K-Means and Hierarchical Clustering algorithms Automatedmodel-based tissue classification of MR images of the brain,
and the results obtained showcase that the developed model has IEEE Trans. Med. Imag., 18(10): 897–908, Oct. 1999.
better segmentation accuracy. Effective segmentation helps in [12] Dugas-Phocion, M. Á. González Ballester, G. Malandain, C. Lebrun and
efficient identification of the damaged tissues much more N. Ayache: Improved EM-based tissue segmentation andpartial volume
effect quantification in multi-sequence brain MRI, Int. Conf. Med.
effectively Therefore, the proposed method will be very much Image Comput. Comput. Assist. Int. (MICCAI), 2004, pp. 26–33.
useful in diagnosing the diseases like acoustic neuroma, [13] K. Van Leemput, F. Maes, D. Vandermeulen, and P. Suetens: A
Alzheimer’s, Parkinson’s etc. more accurately. unifying framework for partial volume segmentation of brainMR
Images, IEEE Trans. Med. Imag., 22(1):105–119, Jan. 2003.
IX. FUTURE SCOPE [14] M. Prastawa, E. Bullitt, S. Ho, and G. Gerig: Robust estimation forbrain
A methodology is presented for analyzing the brain images tumor segmentation, Int. Conf. Med. Image Comput. Comput.Assist.
Inter (MICCAI), 2003, pp. 530–537.
based on Truncated Gaussian Mixture models. However, to
have a more precise segmentation, it is needed to consider the [15] Lee TI, Maximum likelihood estimation for multivariate skew normal
mixture models, Journal of Multivariate Analysis 100 (2009) 257–265,
other features of the images also, which may include the shape, 2009.
size, orientation and texture. Therefore, multivariate features [16] Cabral CS, Lachos VH, Prates MO, Multivariate mixture modeling
should be considered to have a more detailed and effective using skew-normal independent distributions. Computational Statistics
analysis, further work is to be projected in this direction. and Data Analysis 56:126–142, 2012.
REFERENCES [17] Lee TI, Robust mixture modeling using multivariate skew distributions,
stat comp (2010) 20:343-356, DOI 10.1007/s I 1222-009-9128-9, 2010.
[1] Vasant Manohar and Yuhua Gu: MRI Segmentation Using Fuzzy C-
Means and Finite Gaussian Mixture Model, Digital Image Processing – [18] Lee S, McLachlan GJ, On the fitting of mixtures of multivariate skew t-
CAP5400, 2008. distributions via the EM algorithm. arXiv:11094706 [statME], 2011.
[2] Z Y Shan, G H Yue and J Z Liu: Automated Histogram-Based Brain [19] Vrbik I, McNicholas PD, Analytic calculations for the EM algorithm for
Segmentation in T1-Weighted Three-Dimensional Magnetic Resonance multivariate skew t-mixture models. Statistics and Probability Letters,
Head Images, NeuroImage v.17, pp. 1587-1598, 2002. 32 Sharon X. Lee, Geoffrey J. McLachlan 82:1169–1174, 2012.
[3] G V S Raj Kumar, K Srinivasa Rao, and P Srinivasa Rao: Image [20] Lin TI, Ho HJ, Kee CR, Flexible mixture modelling using the
Segmentation Method Based on Finite Doubly Truncated Bivariate multivariate skew-t-normal distribution. Statistics and Computing DOI
Gaussian Mixture Model with Hierarchical Clustering, International 10.1007/s11222-013-9386-4, 2013.
Journal of Computer Science Issues, 8(4-2):151-159, July, 2011. [21] Karlis D, Xekalaki E, Choosing initial values for the EM algorithm for
[4] Nagesh Vadaparthi, Srinivas Yerramalle, and Suresh Varma.P: finite mixtures. Computational Statistics & Data Analysis 41:577–590,
Unsupervised Medical Image Segmentation on Brain MRI images using 2003
Skew Gaussian Distribution, IEEE – ICRTIT 2011, pp.1293 – 1297. [22] Franczak BC, Browne RP, McNicholas PD, Mixtures of shifted
[5] Nagesh Vadaparthi, Srinivas Yerramalle, and Suresh Varma.P: asymmetric laplace distributions. arXiv:12071727 [statME], 2012.
Unsupervised Medical Image Classification based on Skew Gaussian [23] McNeil AJ, Frey R, Embrechts P, Quantitative risk management :
Mixture Model and Hierarchical Clustering Algorithm” in CCIS of concepts, techniques and tools. Princeton University Press, USA, 1997.
Springer-Links, Volume – 203, pp. 65-74, 2011. [24] Embrechts, P., Kl¨uppelberg, C., and Mikosch, T., Modelling Extremal
[6] Nagesh Vadaparthi, Srinivas Yerramalle, and Suresh Varma.P: Events for Insurance and Finance. Springer-Verlag, New York, 1997.
Segmentation of Brain MR Images based on Finite Skew Gaussian [25] Burnecki, K., Misiorek, A., and Weron, R., Loss distributions. In
Mixture Model with Fuzzy C-Means Clustering and EM Algorithm”, Statistical Tools for Finance and Insurance. Cizek, P., H¨ardle, W. K.,
International Journal of Computer Applications, 28(10):18-26, August and Weron, R. Eds. Springer-Verlag.2010.
2011. [26] Sylvain Bouix et al.: Evaluating Brain Tissue Classifiers without a
[7] Nagesh Vadaparthi, Srinivas Yerramalle, and Suresh Varma.P: An ground truth, Journal of NeuroImage (ELSEVIER) - 36, pp. 1207 –
Efficient Approach for Medical Image Segmentation Based on 1224, 2007.
Truncated Skew Gaussian Mixture Model Using K-Means Algorithm”, [27] Ahmet M. Eskicioglu and Paul S. Fisher: Image Quality Measures and
International Journal of Computer Science and Telecommunications, Their Performance, IEEE Transactions on Communications,
(ISSN 2047-3338), 2(6):79-86, August 2011. 43(12):2959 – 2965, Dec.1995.

197 | P a g e
www.ijacsa.thesai.org