A New Hybrid Method For The
A New Hybrid Method For The
A New Hybrid Method For The
4, August 2014
DOI : 10.5121/sipij.2014.5408 77
A NEW HYBRID METHOD FOR THE
SEGMENTATION OF THE BRAIN MRIS
Mohamed Zaki Abderrezak, Mouatez billah Chibane, Prof. Karim Mansour
Laboratory of Study of Electronic Materials for Medical applications
University Constantine 1 - Road of Ain El Bey - Constantine - Algeria.
ABSTRACT
The magnetic resonance imaging is a method which has undeniable qualities of contrast and tissue
characterization, presenting an interest in the follow-up of various pathologies such as the multiple
sclerosis. In this work, a new method of hybrid segmentation is presented and applied to Brain MRIs. The
extraction of the image of the brain is pretreated with the Non Local Means filter. A theoretical approach is
proposed; finally the last section is organized around an experimental part allowing the study of the
behavior of our model on textured images. In the aim to validate our model, different segmentations were
down on pathological Brain MRI, the obtained results have been compared to the results obtained by
another models. This results show the effectiveness and the robustness of the suggested approach.
KEYWORDS
Hybrid Segmentation, Petreatment, Cerebral .MRI, Multiple Sclerosis, tumors.
1. INTRODUCTION.
The segmentation plays a big role in the image processing. It is carried out before the stages of
analysis and catch of the decisions in several image treatments of, such as the detection of the
cerebral anomalies in the magnetic resonance images.
During these last years several methods of segmentation were developed [1,2,3,4,5,6]. among
them active contours (ACM) which was applied to various types of medical imaging modalities.
Generally in the literature, we distinguishe two types of (ACM): model based contour [1,2,24]
and model based area [3,4,5,6]. A good review of the principal models of active contours can be
found in the two following references [7,8].
To locate the areas borders in the segmentation based contour, we can use the gradient
information. This method allows the detection of the transitions between the various areas of the
image [1,2,3]. The detectors of contours used can be simple, like the operators of Sobel and
Roberts [9], or more complexes such as the operator of Canny [11]. The results of this
segmentation are the candidates of the limits of the image objects. However, they can present
false detections because its models remain sensitive to the noises.
The principle of the segmentation based area is to take into account some descriptors of areas,
such as the intensity, the color, the texture or the movement [13]. The function to be minimized in
this case, is written like an integral on the area and not on contour. However, the areas obtained
do not correspond in all cases, with the objects represented in the image. The limits of the
obtained areas are usually vague and do not coincide exactly with the bornes of the objects of the
image (Kermad and Chehdi, 2002) [14].
Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.4, August 2014
78
One of the most used methods of segmentations based area in imaging, is the model of Chan and
Vese [3]. It had much success on binary images with the assumption that each area in the image,
is homogeneous. While on medical images, the same area is found with various degrees of
homogeneities, from where multiple modifications made to the model of Chan and Vese in order
to adapt it to different type of medical imaging modalities [15].
Another approach of segmentation called hybrid segmentation, combining the advantages of the
two preceding ones, was developed [16,17]. The method based area, aims at the detection of the
ROI (limits of the regions of interest), and uses the segmentation based contour to determine the
limits of these areas. The algorithms combining the techniques of segmentation based on the areas
and those based on contours take advantage of the complementary of the information on the area
and contour [18]. Each step is executed iteratively allowing the other to improve. Yu et al. [19]
fused methods by set of level (implicit deformable models) with Bayesian techniques to segment
the cortex. Yang and Kruggel [20] combine Bayesian approaches with Watershed algorithms to
segment the cortical sulci.
Our work concerns the development of a new hybrid approach starting from a geodetic active
contour. This approach is tested on cerebral images obtained by magnetic resonance. In our
model the gradient is replaced by the GVC (Gradient Vector Convolution) to widen the card of
capture.
2. MATERIAL AND METHOD.
The problem of the detection of the artifacts in the medical images hasnt a single solution. It is
necessary however, to choose most adequate for the images obtained by MRI which are affected
by a random noise which can deteriorate the segmentation. Before carrying out detection, we
have developed a new algorithm to eliminate all the parts around the brain which present little
interest, even not interest for the physician (example skin of the head).
Step 1: to calculate the variance in various scales and various directions with an aim of
extracting a feature set of the initial image knowing that:
varx =
x
x
n
Step 2: to calculate the gradient and the coefficient of conduction.
Step 3: to use the algorithm of Otsu to obtain a binary image, and to benefit from the level of
gray of the cranium, which is darker than the cerebrospinal fluid which surrounds the
brain.
Step 4: To carry out a morphological opening using a disc of ray 5 pixels on the binary image,
which makes it possible to eliminate the smaller related objects.
Step 5: To fill the parts isolated from the brain not eliminated at the time of the preceding
stage.
Step 6: To superimpose the two images to make the extraction.
Step 7: Final image.
To eliminate the artifacts from the image, we apply a Non Local Means filter, which proved its
efficiency on the MRI [10]. Being given the image using method NLM the value filtered in a
point is calculated like a weighted average of all the pixels of the image given by:
NLMux = wx, y
. uy
0 wx, y 1, wx, y
= 1 (1)
Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.4, August 2014
79
) , ( y x w represents the NL-means weights according to the similarity between the values of the
vicinities and are written:
wx, y =
1
Zx
e
Zx = e
(2)
a parameter which allows the control of the filter, represents a constant of standardization, and d
the Euclidean distance which is written:
, =
avec
[1.5] (3)
The obtained results (figure 2) show the effectiveness of our algorithm to carry out the extraction
of the brain image in the aim to prepare it to the segmentation.
Initial Image final Image
Figure 2. Pretreatment of a MRI with our algorithm
When the pretreatment is finished, we carry out the segmentation by using the model of
geodetic contours proposed by Casselles et al. [21] and given by the following expression
,,
=
. + . +
(5)
Where 0 , 0,
> 0
represent positive constants.
Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.4, August 2014
80
epresent two constants which define the approximation of the intensity respectively
inside and outside contour C.
After minimization of the function (5) Chan and Vese obtain the following formula:
= . [.
] (6)
Where =
is the Dirac function.
represents the function of curvature, calculated with the following equation:
=
||
(7)
C
=
.
(8)
The term
I C
I C
= ||. [.
] (9)
We suppose that
and by summoning it with the geodetic model of the expression
(4), we obtain the following function:
= ||. +
+ . (10)
is a constant which ensures the proportionality between the term of the forces of gradient and
the term of the forces of area. g represents the function of stop which is calculated by the
following expression :
g =
with 1 (11)
Where
is the Gaussian core. The second term of the equation was replaced by term
representing the GVC (Gradient Vector Convolution) [23], to regulate the problem of the
dependence to the initialization of contour or to obtain a contour which is able to be driven to the
border of the regions of interest.
The approximation of the equation (10) is then written in the following form:
+t. [|| g. u. k + c
u +
+ w
. w
] (12)
c
, c
. w
) represent GVC.
Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.4, August 2014
81
To carry out the segmentation, we have used the following algorithm:
Step 1 :
, = ,
Step 6 : Calculation of the gradient g:
For i=1 to ln
For j=1 to Cn do
|, | =
1
1 + |,
, |
Step 7 : Calculation of the term of curvature K:
For i=1 to ln
For j=1 to Cn do
, =
,
|,|
Step 8 : Calculation of
2 1
, c c one using the function (8).
Step 9 : Calculation of
w w
g
. the Gradient Vector Convolution.
Step 10 : Calculation of
t+1
(i,j)
+ . , . +
, +
2
+
,
Step 11 : Outputs,
,