19-06-2014, 02:52 PM
EXTRACTION OF BRAIN TUMORS FROM MRI IMAGES WITH
ARTIFICIAL BEE COLONY BASED SEGMENTATION METHODOLOGY
[attachment=65065]
Abstract
Image segmentation plays significant role in medical
applications to extract or detect suspicious regions. In this
paper, a new image segmentation methodology based on
artificial bee colony algorithm (ABC) is proposed to extract
brain tumors from magnetic reasoning imaging (MRI), one
of the most useful tools used for diagnosing and treating
medical cases. The proposed methodology comprises three
phases: enhancement of the original MRI image (preprocessing),
segmentation with the ABC based image
clustering method (processing), and extraction of brain
tumors (post-processing). The proposed methodology is
compared and analyzed on totally 9 MRI images shooting in
different positions from a patient with the methodologies
based on K-means, Fuzzy C-means and genetic algorithms.
It is observed from the experimental studies that the
segmentation process with the ABC algorithm obtains both
visually and numerically best results.
1. Introduction
Segmentation is one of the most significant requirements in
the analysis of medical images. However, the complicated
structures of the inside of the human body cause problems in
segmentation i.e. prescribing appropriate therapy. Magnetic
resonance imaging (MRI) is a technique primarily used in
medical conditions to get high quality images of organs, soft
tissues, bone and virtually all other internal human body
structures. In other words, “MRI possesses good contrast
resolution for different tissues and has advantages over
computerized tomography (CT) for brain tissues due to its
superior contrast properties [1]”. On account of those
advantages, MRI images have become a basic source of medical
image segmentation, especially brain segmentation. Brain MRI
segmentation is mainly applied to the following fields [2]: 1)
automatic or semiautomatic diagnosis of regions to be treated
prior to the surgery, 2) diagnosis of tumors before and after
surgical intervention for response assessment, and 3) tissue
classification. This paper particularly concerns with the second
field.
Traditionally, segmentation of brain MRI images is
processed manually by radiologists. However, manual
segmentation is high time consumption and cause unavoidable
mistakes. To sort out these problems, researchers proposed
various segmentation methods based on thresholding [3],
boundary detection [4], region growing [5], and clustering [6].
Thresholding, the process of separating the regions regarding to
pixel intensity and color, is one of the simplest methods using
for segmentation. Although thresholding methods can be
efficiently applied to distinguish the objects from background
when the histogram of objects and background is apparently
distributed, these methods ignore all the spatial information of
an image and cannot deal with the noise. Therefore they cannot
reach adequate performance in segmentation of brain MRI
images. Boundary detection methods try to find rapid pixel
change between the regions at the boundary. That is achieved by
gradient operator, e.g., Sobel and Prewitt operators. These
methods are very dependent to the noise and describing edge
pixel members as boundaries of regions is a challenging process.
Region growing methods rely on the neighboring pixels having
similar values within the region. To achieve the satisfactory
results in region growing methods, the regions must be
homogenously appeared in an image and it requires a user
intervention to select the seed candidates and to determine the
criteria of region growth [5]. Image clustering which is the
process of partitioning patterns such as pixels into groups or
clusters using similarity criterion is mostly used method in brain
MRI segmentation. The most well-known clustering methods
are K-means [7]. Fuzzy C-Means (FCM) [8]. K-means, one of
the simplest clustering methods, is based on the similarity
criterion such that Euclidean distance between the patterns and
the cluster centroids and FCM, the fuzzy version of K-means, is
based on updating the memberships over the cluster centroids.
Although, they are very popular in brain MRI segmentation,
they are very dependent to initial conditions i.e. the centroids or
memberships should be determined in an efficient way to
improve the performance of clustering quality. In addition, they
can adversely affect from noise i.e. that is not convenient for
FCM since brain MRI images include considerable unknown
and uncertain noise [9]. To minimize the drawbacks of FCM,
Shen et.al. [1] proposed an improved FCM considering the
difference between the neighboring pixels in the image and the
relative locations of neighboring pixels. That method therefore
does not only consider pixel intensities, but also consider
neighboring the pixel intensities and locations. The degree of
neighboring attractions is defined by the parameters determined
by simple artificial neural networks (ANN). Hence, the
performance of the proposed method relies on these parameters.
In this paper, artificial bee colony based image segmentation
methodology proposed to extract brain tumors from MRI
images. The proposed methodology is constructed on three
stages. In first stage, original image is enhanced and eliminated
from noise by 2D 3x3 median filter. In second stage, artificial
bee colony based image clustering method is applied to the
enhanced image. In the last stage, the segmented image is
converted into binary image using thresholding and then
connected component labeling method is employed to extract
brain tumor.
The organization of the paper is as follows. In Section 2, the
artificial bee colony algorithm is presented. In Section 3, the
proposed segmentation methodology is introduced and the
algorithms applied for analysis are expressed. The experimental
studies are demonstrated to compare the proposed algorithm
with the others in Section 4 and finally, the paper is concluded.
Experimental Studies
To verify the effectiveness of the proposed methodology,
nine MRI images called as mri70-mri150 (256x256 pixels) from
a patient are chosen, and the K-means, FCM and GA methods
are employed. The methods are evaluated in terms of both
numerical and visual results. Numerically, the methods are
analyzed by calculating CS measurement [18] and the Turi’s VI
index [19] which are the clustering validity indexes to measure
the clustering quality, defined by Eq. 12 and 13, respectively.
Visually, the methods are compared by using the output images
of extracted tumors, illustrated in Fig. 4. The visual results of
GA is not presented in Fig 4 since it is suggested from the visual
experiments that it cannot be possibly to illustrate apparent
differences between the output images of the GA and ABC
algorithms.
Conclusions
In this paper, the artificial bee colony based image
segmentation methodology is proposed to extract the region of
tumors from MRI images. The proposed methodology is
analyzed and compared with the K-means, FCM, and GA based
image segmentation methodologies. The both visual and
numerical results indicate that the ABC based image
segmentation methodology outperforms the others and can be
efficiently used in tumor segmentation of MRI images. The
future work is to improve the performance of the proposed
clustering methodology.
ARTIFICIAL BEE COLONY BASED SEGMENTATION METHODOLOGY
[attachment=65065]
Abstract
Image segmentation plays significant role in medical
applications to extract or detect suspicious regions. In this
paper, a new image segmentation methodology based on
artificial bee colony algorithm (ABC) is proposed to extract
brain tumors from magnetic reasoning imaging (MRI), one
of the most useful tools used for diagnosing and treating
medical cases. The proposed methodology comprises three
phases: enhancement of the original MRI image (preprocessing),
segmentation with the ABC based image
clustering method (processing), and extraction of brain
tumors (post-processing). The proposed methodology is
compared and analyzed on totally 9 MRI images shooting in
different positions from a patient with the methodologies
based on K-means, Fuzzy C-means and genetic algorithms.
It is observed from the experimental studies that the
segmentation process with the ABC algorithm obtains both
visually and numerically best results.
1. Introduction
Segmentation is one of the most significant requirements in
the analysis of medical images. However, the complicated
structures of the inside of the human body cause problems in
segmentation i.e. prescribing appropriate therapy. Magnetic
resonance imaging (MRI) is a technique primarily used in
medical conditions to get high quality images of organs, soft
tissues, bone and virtually all other internal human body
structures. In other words, “MRI possesses good contrast
resolution for different tissues and has advantages over
computerized tomography (CT) for brain tissues due to its
superior contrast properties [1]”. On account of those
advantages, MRI images have become a basic source of medical
image segmentation, especially brain segmentation. Brain MRI
segmentation is mainly applied to the following fields [2]: 1)
automatic or semiautomatic diagnosis of regions to be treated
prior to the surgery, 2) diagnosis of tumors before and after
surgical intervention for response assessment, and 3) tissue
classification. This paper particularly concerns with the second
field.
Traditionally, segmentation of brain MRI images is
processed manually by radiologists. However, manual
segmentation is high time consumption and cause unavoidable
mistakes. To sort out these problems, researchers proposed
various segmentation methods based on thresholding [3],
boundary detection [4], region growing [5], and clustering [6].
Thresholding, the process of separating the regions regarding to
pixel intensity and color, is one of the simplest methods using
for segmentation. Although thresholding methods can be
efficiently applied to distinguish the objects from background
when the histogram of objects and background is apparently
distributed, these methods ignore all the spatial information of
an image and cannot deal with the noise. Therefore they cannot
reach adequate performance in segmentation of brain MRI
images. Boundary detection methods try to find rapid pixel
change between the regions at the boundary. That is achieved by
gradient operator, e.g., Sobel and Prewitt operators. These
methods are very dependent to the noise and describing edge
pixel members as boundaries of regions is a challenging process.
Region growing methods rely on the neighboring pixels having
similar values within the region. To achieve the satisfactory
results in region growing methods, the regions must be
homogenously appeared in an image and it requires a user
intervention to select the seed candidates and to determine the
criteria of region growth [5]. Image clustering which is the
process of partitioning patterns such as pixels into groups or
clusters using similarity criterion is mostly used method in brain
MRI segmentation. The most well-known clustering methods
are K-means [7]. Fuzzy C-Means (FCM) [8]. K-means, one of
the simplest clustering methods, is based on the similarity
criterion such that Euclidean distance between the patterns and
the cluster centroids and FCM, the fuzzy version of K-means, is
based on updating the memberships over the cluster centroids.
Although, they are very popular in brain MRI segmentation,
they are very dependent to initial conditions i.e. the centroids or
memberships should be determined in an efficient way to
improve the performance of clustering quality. In addition, they
can adversely affect from noise i.e. that is not convenient for
FCM since brain MRI images include considerable unknown
and uncertain noise [9]. To minimize the drawbacks of FCM,
Shen et.al. [1] proposed an improved FCM considering the
difference between the neighboring pixels in the image and the
relative locations of neighboring pixels. That method therefore
does not only consider pixel intensities, but also consider
neighboring the pixel intensities and locations. The degree of
neighboring attractions is defined by the parameters determined
by simple artificial neural networks (ANN). Hence, the
performance of the proposed method relies on these parameters.
In this paper, artificial bee colony based image segmentation
methodology proposed to extract brain tumors from MRI
images. The proposed methodology is constructed on three
stages. In first stage, original image is enhanced and eliminated
from noise by 2D 3x3 median filter. In second stage, artificial
bee colony based image clustering method is applied to the
enhanced image. In the last stage, the segmented image is
converted into binary image using thresholding and then
connected component labeling method is employed to extract
brain tumor.
The organization of the paper is as follows. In Section 2, the
artificial bee colony algorithm is presented. In Section 3, the
proposed segmentation methodology is introduced and the
algorithms applied for analysis are expressed. The experimental
studies are demonstrated to compare the proposed algorithm
with the others in Section 4 and finally, the paper is concluded.
Experimental Studies
To verify the effectiveness of the proposed methodology,
nine MRI images called as mri70-mri150 (256x256 pixels) from
a patient are chosen, and the K-means, FCM and GA methods
are employed. The methods are evaluated in terms of both
numerical and visual results. Numerically, the methods are
analyzed by calculating CS measurement [18] and the Turi’s VI
index [19] which are the clustering validity indexes to measure
the clustering quality, defined by Eq. 12 and 13, respectively.
Visually, the methods are compared by using the output images
of extracted tumors, illustrated in Fig. 4. The visual results of
GA is not presented in Fig 4 since it is suggested from the visual
experiments that it cannot be possibly to illustrate apparent
differences between the output images of the GA and ABC
algorithms.
Conclusions
In this paper, the artificial bee colony based image
segmentation methodology is proposed to extract the region of
tumors from MRI images. The proposed methodology is
analyzed and compared with the K-means, FCM, and GA based
image segmentation methodologies. The both visual and
numerical results indicate that the ABC based image
segmentation methodology outperforms the others and can be
efficiently used in tumor segmentation of MRI images. The
future work is to improve the performance of the proposed
clustering methodology.