Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
International Journal of Academic Research in Computer Engineering
Print ISSN:2476-7638 and Online ISSN: 2538-2411 Vol. 1, No. 2, Pages. 57-67, November 2016
57
International Journal of Academic Research in
Computer Engineering
Volume 1, Number 2, Pages 57-67, November 2016
www.ijarce.org
IJARCE
A New Hybrid of Active Contour and Multiphase Level Set Models for
Segmenting of Medical Images: A Case Study
Siamak Abdehzadeh1*, Naser Sharifi2, Barfab Wafaee3, Majid Habibi4, and Mohammad
Hosntalab5
1. Department of Computer, Boukan Branch, Islamic Azad University, Boukan, Iran.
2. Department of Computer, Boukan Branch, Islamic Azad University, Boukan, Iran.
3. Department of Computer, Baneh Branch, Islamic Azad University, Baneh, Iran.
4. Department of Computer, Tabriz Branch, Islamic Azad University, Tabriz, Iran.
5. Faculty of Engineering, Science and Research Branch, Islamic Azad University, Tehran, Iran.
Receive Date 2016.10.02; Accepted Date: 2016.11.15, Published Date: 2016.11.15
*Corresponding Author: S.Abdehzadeh ([email protected])
Abstract
Medical images segmentation due to the increasing volume of this images is difficult and in accessible to
human. with development of image processing we can use computers to help. Segmentation of the early stages
of image processing are very well regarded. In this paper, a combination method based on level set and active
contour models have been proposed to achieve more accurate image in image processing. The selected dataset
contains different slices that in this article 8 slices have been selected for testing. The proposed algorithm was
applied on these slices and it has been compared with previous methods. The results have been tested on MRI
images of the brain and this results show that the proposed method is better than other methods proposed
before.
Keywords: Medical Image Segmentation; Level Set Model; Active Contour; Brain MRI Image.
1. Introduction
Segmentation is a fundamental technique in image
processing, which is still being developed and
completed, Level set of the practical methods for
selecting region of interest. Enhance the accuracy
of segmentation methods in image processing is a
Specified target. Besides adding new techniques
such as multi-resolution and multi-phase can be
lead to achieve better results. In this study, the
addition of new methods noted, to Level set and
adjust its parameters lead to achieve better results.
This research is also being evaluated as a new
method by researchers [1,2,3]. Due to recent
advances, segmentation can be used directly in
treatment methods, Thus increasing the accuracy is
very clear [3]. Segmentation is done naturally in
the human visual system. The aim of threshold
algorithms, is to finding an ideal threshold value
for better segmenting of images. The threshold can
be adjusted either manually or automatically.
Manually choosing the threshold, one must
consider whether the threshold value is correct or
not. This process is time consuming. In choosing of
T, errors may occur and cause problems in image
analysis later. For this reason, many methods of
automatic threshold are introduced [4, 5, 6]. The
most popular active contour method introduced in
[7]. Kass named this method snakes. Because
during evolution, the contour moving toward
object that like a snake. According to the estimates
of the boundary of an object in an image, as initial
contour, snakes locate the actual border. In [8]
focus on reducing dependence on initial conditions
by defining a new external energy for improving
snake algorithm. Cohen et. al. in [9] proposed a
S. Abdehzadeh et.al. / International Journal of Academic Research in Computer Engineering
58
new snake algorithm as balloon snake and added
second external force which is deform the contour
in the normal direction to the outer direction
(inflation) or inner direction (condensation). First,
the level set method was introduced in [10], then
by [11] changed. To getting moving front in a wide
range of issues, the level set method, is a powerful
numerical selection. The areas where the level set
method can be used are image processing,
computer vision and graphics. As noted by [12] It
is said, an implicit representation of the super
planes data, Set of PDE's which controls the
movement of planes manner, and related numerical
methods for implementing in computers make
elements of classical level set method.
2. Material and Methods
2.1. Multi-Stage Level Set Formulation and
Active Contour Model
We use multi-level set functions i, i = 1... n to
represent the regions Ωi, i = 1... N with N = 2n
as in [13]. That we consider Mi(φN(.))as
characteristic function of region Ωi, N (.)φ is a
function of φi, i=1… n, and called statistical
variable Multi-fuzzy level set (SVMLS) we can
present formulation of SVMLS in Equation (1) on
the basis of characteristic function [14].
(y) ( (y))
N
nSVMLSd M dy
i iN
(1)
Where d is equal to Equation (2):
(2)
We minimize the energy for four fuzzy level show
in the Equation (3).
𝑀1 = 𝐻(φ1)𝐻(φ2)
𝑀2 = 𝐻(φ1)(1 − 𝐻(φ2))
𝑀3 = (1 − 𝐻(φ1))𝐻(φ2)
𝑀4 = (1 − 𝐻(φ1))(1 − 𝐻(φ2)
(3)
In which H (φ) is heavy side function we often in
practice, use the regular version in Equation (4).
𝐻𝜀 = 1
2 [1 +
2
𝜋 arctan (
φ
휀)]
(4)
But the process of minimizing the energy includes
maintaining a variable constant, and changing the
other variable, we try to minimize energy that
Equation (5) and Equation (6) and Equation (7) to
calculate b, and 𝜎𝑖 , 𝑖 = 1,2,3,4 are listed:
𝑏 =
∑ (𝐾𝑝 ∗ (𝐼𝑀𝑖(φ4)) ∗𝑐𝑖𝜎𝑖2)
4𝑖=1
∑ (𝐾𝑝 ∗ (𝑀𝑖(φ4)) ∗𝑐𝑖𝜎𝑖2
4𝑖=1 )
(5)
𝐶𝑖 = ∫(𝐾𝑝 ∗ 𝑏) 𝐼𝑀𝑖(φ4)𝑑𝑦
∫(𝐾𝑝 ∗ 𝑏2) 𝑀𝑖(φ4)𝑑𝑦
(6)
𝜎𝑖 = √∬𝐾𝑝 (𝑦𝑥)((𝐼(𝑦) − 𝑏(𝑥)𝑐𝑗)2𝑀𝑖(φ4(𝑦))𝑑𝑦𝑑𝑥
∬𝐾𝑝 (𝑦𝑥)𝑀𝑖(φ4(𝑦))𝑑𝑦𝑑𝑥 (7)
Thus, by minimizing the energy function 𝐸φ4 𝛼𝑆𝑉𝑀𝐿𝑆
for φ1 and 𝜑2. We get gradient decline related to
that as Equation (8). 𝜕Φ1𝜕T
= −[(𝑑1 − 𝑑2 − 𝑑3 + 𝑑4)𝐻(φ2) + 𝑑2 − 𝑑4]𝜕(Φ1) (8) 𝜕Φ2
𝜕T= −[(𝑑1 − 𝑑2 − 𝑑3 + 𝑑4)𝐻(φ1) + 𝑑2 − 𝑑4]𝜕(Φ2)
Where δ (φ) is direct function that follows from
Equation (9).
𝛿𝜀(Φ) =1
𝜋∗
휀
휀2 + Φ2 (9)
But we must ensure that φ remained normalized.
That their vectors size is 1. This is done by
multiplying by a kernel function. The general trend
of the algorithm is as follows:
1. φ values, are randomly initialize.
2. φ values are considered constant and
update b, c and ϭ.
3. b, c and ϭ are considered constant and
update φ.
4. If convergence is not reached, go to 2.
We consider the kernel simple with a small
dimension (3*3). The Radius of ρ is also
considered 4.5. We keep the output of the model
SVMLS in order to combine it with active contour
model.
2.2. Level Set Method Two methods of statistical and variable Multi-fuzzy level set (SVMLS) and weighted K-means vibrational level set (WKVLS) which are from main methods of the level sets are shown in Figures 1 and 2 and then by combining this approach with active contour model we will see its results with SVMLS at the tests and results [14].
Figure 1. SVMLS Method [14]
S. Abdehzadeh et.al. / International Journal of Academic Research in Computer Engineering
59
Figure 2. WKVLS Method [14]
In the Figure (1) and Figure (2) from left to right,
we see the initialization value of the algorithm, the
second figure is an estimate of bias, the
segmentation result is shown in the third figure and
the fourth one is a figure in which the bias is
corrected. In general, it is observed that SVMLS
method compared to WKVLS is better segmented
the image (partly shown with the red mark).
2.3 Dataset
Selected image from dataset are of the size
181*217*181 pixel, image model T1, protocol of
the type ICBM, phantoms name is normal, Slice
thickness is 1 mm, noise in all three datasets is
equal to 0%, 3% and 9% respectively. Non-
uniform intensity in all three datasets is 20%. The
color scheme is gray in all three, all dataset is equal
except that their noise percentage varies, all
datasets saved in MINC format [15] and from each
dataset we choose 8 slice with numbers
86،87،88،89،90،91،92،95.
Comprised of brain images of 19 people with
dimensions of 136*189*157 pixels. Database sored
in .nii format and algorithms have tested on 72
slices of 19 selected brain images [15].
3. Proposed Method
The combination of SVMLS method with active
contour the following have explained. Overall
flowchart for segmentation methods and
calculation errors are shows in Figure (3).
Figure 3. Overall flowchart for segmentation methods
and calculation errors
One method of segmentation is using active contour
models. Here we are going to use a combination of
active contour models and methods obtained from
SVMLS and WKVLS to achieve better accuracy.
The study and use of active contour, will lead to
desirable results in segmentation. This approach is
based on the using variable contours which move
under the effect of this force and to tracking
boundaries and move are used. The idea of using a
flexible model to select specific features within an
image for the first time in [16] has been introduced,
nothing had not been done in this area. The goal is
to find an equation that can obtain a picture’s
contours .In other words, there must be a curve that
its boundaries segments the object. Since each
model requires a lot of steps to get good accuracy in
the segmentation, each method has its own time
needed to reach result. But this can be done with
fewer repetitions and combine them to achieve
better accuracy in less time. We use Chan-Vese
algorithm [19,20], An energy function is defined.
Aim will be to minimize the energy function and
minimizing the levels set Φ will do the
segmentation. In general, the energy function is
shown in the Equation (10).
S. Abdehzadeh et.al. / International Journal of Academic Research in Computer Engineering
60
F(Φ)=𝜇(∫ |∇𝐻(Φ)|𝑑𝑥Ω)𝑝 +
𝑣 ∫ 𝐻(Φ)𝑑𝑥 + 𝜆 ∫ |𝐼 − 𝑐1|2
ΩΩ𝐻(Φ)𝑑𝑥 +
𝜆2 ∫ |𝐼 − 𝑐2|2 (1 − 𝐻(Φ)𝑑𝑥
Ω
(10)
The parameters can be adjusted by the user. The
general case of Mumford-Shah function. In The
Mumford-Shah we set our parameters as P=1, v=0,
λ1=λ2=1 [17, 13]. H and I are defined as in the
previous step. C1 and C2, respectively, at the time
that ф is larger or smaller than zero it's defined as
Equation (11).
1 2
. ( ) .(1 H( ),
( ) (1 ( ))
I H dxdy I dxdyc c
H dxdy H dxdy
(11)
Using the Euler-Lagrange and gradient descent for
the level set function and the derivative we
minimize energy function as (12):
1
2 2
1 1 2 2
( )( )
( ) ( )
p
t
p div
I c p I c
(12)
After the discretization and linearization of
differential equations of minimizing the energy
function that is shown in Equation (13). 1
. ,
1
. 2
1
.
2 2 2 2
. . .
1
.
2 2 2 2
1. .
2
. 1 . 1 2
( ) ( . ( ) )
( ) / ( 1 1) / (2 )
( 1. ) / (2 ) ( ) / (2 )
( )( ( ( )) (
n n
i j i j
n n p
h i j
x n
i jx
x n n n
i j i j i j
x n
i jy
x n y n
i i j i j
n n
h i j i j i
t
p Lh
h h
j h h
I c I
2
. 2 ( )) .n
j c
(13)
δ is softened state of the delta function according to Equation (14).
2 2
1( )h
hx
h x
(14)
L is the length of the levels set in zero step and according to Equation (15) we have.
( ) ( )n n n
hL dxdy
(15)
We consider the Equation (16) constants for the
Equation (15).
12 2/4
1. . , 1 . 1
22 2/4
1. . 1, 1 1. 1
32/4 2
1. 1. , 1 .
42/4 2
1. 1. , .
1
( ) ( )
1
( ) ( )
1
( ) ( )
1
( 1) ( 1)
n n n n
i j i j i j i j
n n n n
i j i j i j i j
n n n n
i j i j i j i j
n n n n
i j i j i j i j
C
C
C
C
(16)
Using the Equation (16) definition finally we
achieve the Equation (17).
1 2 3 4
2 41
1 1
. .
1 1
. .
1 1 1 1
1. 1. 3 , 1 . 1
.
2 2
1 . 1 2 . 2
1 ( ) ( . ( ) ( )
( ) ( . ( ) )
( )
( )
( ( )) ( ( ))
n n n p
i j h i j
n n n p
i j h i j
n n n n
i j i j i j i j
n
h i j
n n
i j i j
tp L C c c c
h
tp L
h
C c c c
t
I c I c
(17)
The algorithm presented in Equation (17) is Chan-Vese algorithm [18,19,20] for segmentation. To increase the accuracy of the segmentation, output of the model that is an image with the same size as the original image is multiplied Pixel by pixel to the result of the previous step that is also with the same size as original image, to obtain the final optimal results.
4. Evaluation and Results
In this section, we discuss and compare the results
between the proposed approach and methods of
SVMLS and WKVLS. The method will be tested
on a phantom circle. Figure 4 consists of six circles
with different illumination intensities of 60, 90,
120, 130, 200 160 and Gaussian noise with given
variance and mean intensity of the class is also
added. Figure 4 shows the phantom simulation of
the results. And Figure 5 shows diagram without
phantom noise.
S. Abdehzadeh et.al. / International Journal of Academic Research in Computer Engineering
61
Figure 4. Circle phantom simulation
Figure 5. Diagram Without Noise
According to the diagram, the green line shows
original image and red line shows the result of the
proposed method, and the black lines show
SVMLS method. It is clear that the proposed
algorithm has followed the behavior of the image
very well. Figure 6 shows phantom simulation with
noise and Figure 7 shows diagram phantom noise.
Figure 6. Diagram with phantom noise
Figure 7. Intensities of pixels of the image and noise
simulation phantom.
The implementation results are shown below. It
should be noted that the parameters required to
implement the algorithm that can be observed in
Table 1.
Table 1. Basic Parameters of the Proposed Algorithm
Values Parameters Name
4.5 sigma
1 epsilon
.45 time step
7 Iteration_No_of_Algorthim_Runs
1/9*ones(3) Regularizing_Kernel_Filter
ones(4*sigma+1) Initializing_SVMLS_Kernel
4 Seg_Regions_No
4 Max_No_of_Processing_Images
The parameters which are initialized in Table 1 are
from the initial parameters to implement the
algorithms that have been written within the
MATLAB software. At the following, we will do
the results on brain images dataset, which the
selected images contains without noise, noise,
noise 3% and 9% that the obtained results are
represented Figures 8 , 9 and 10.
S. Abdehzadeh et.al. / International Journal of Academic Research in Computer Engineering
62
Figure 8. Segmentation on T1-MRI images without the
noise
Figure 9. Segmentation on T1-MRI image with 3% noise
Figure 10. Segmentation on T1-MRI image with 9%
noise
If we assume, top left the first image and bottom
right the ninth image, Respectively, the first image
is the result of segmentation of all segments, the
next image is sum of all the segments with equal
weight and the next of that is the sum of all
weighted image with same weight and in fact the
output of segmenting algorithm SVMLS. The
fourth image shows the original image, fifth Image,
displays active contour model and the sixth image
illustrates the combinative output and thus the
proposed method. Bottom row from left, shows the
result of masked SVMLS with ideal mask. This
action for result of the proposed method illustrated
in the middle of the bottom row. Finally, in the
lower right hand manually segmentation is
presented. It is clear that the proposed method
compared to SVMLS method is less different than
the manually segmentation method is.
To better illustrate the obtained results and to better
assess the results, we will use the curve of Receiver
operating characteristics (ROC) That the results of
ROC curve on the circle phantom without noise
and taking into consideration the noise are shown
in Table 2 and Figures 11 and 12. The parameters
of the ROC among other parameters used for the
analysis scan be noted these parameters are
including Sensitivity, Specificity, Precision,
Accuracy and Mean error rate. To obtain this
parameter you must convert image to binary, as
well as to obtain the parameter values of the ROC,
for each image, we have two thresholding step. We
have done this for masked result of the proposed
method, SVMLS method and manually
segmentation. And the values of TP, TN, FP and
FN calculated and from the values obtained here,
the five analytical components for the images of
circle simulating phantom and T1-MRI were
calculated. The Image of phantom of circle with
noise and without noise threshold with value of 0.2,
0.3 and 0.4. And results of thresholding for
SVMLS method, proposed method and for
manually segmented image (standard criterion) is
calculated. For the phantom of simulating circle
with noise and without noise, obtained point
calculated in Table 2 and it can be observed there.
S. Abdehzadeh et.al. / International Journal of Academic Research in Computer Engineering
63
Table 2. ROC PARAMETRS CANCULATED BY THE PROPOSED METHOD AND THE METHOD SVMLS FOR
CIRCLE PHANTOM
ROC curve for without noise circle phantom represented in Figure 11 and ROC curve for noisy circle phantom represented in Figure 12.
Figure 11. ROC curve for without noise circle phantom with proposed method and SVMLS method
Figure 12. ROC curve for noisy circle phantom with
proposed method and SVMLS method
In this part, the ROC curve results are done on the
brain images whose results are presented in Tables
3, 4 and 5. We have three T1-MRI dataset of the
brain (without noise, 3% noise, 9% noise) from
Circle phantom Without Noise
SVMLS Method Proposed Method
Mean
error rate
Accurac
y
Precisio
n
Specificit
y
Sensitivit
y
Mean
error rate
Accurac
y
Precisio
n
Specificit
y
Sensitivi
ty Threshold
0.0469 0.9531 0.7196 0.9467 1 0.007
4 0.9926 1 1 0.9385 0.2 Circle
withou
t noise 0.0469 0.9531 0.7196 0.9467 1 0.007
4 0.9926 1 1 0.9385 0.3
Circle phantom 0.001 Noise
SVMLS Method Proposed Method
Mean error rate
Accuracy
Precision
Specificity
Sensitivity
Mean
error
rate
Accuracy
Precision
Specificity
Sensitivity
Threshold
0.0322 0.9678 0.8969 0.9870 0.8270 0.023
4 0.9766 1 1 0.8058 0.2
Circle
0.001
noise
0.0333 0.9667 0.8860 0.9854 0.8295 0.022
9 0.9771 1 1 0.8093 0.3
0.0314 0.9686 0.9040 0.9880 0.8264 0.023
1 0.9769 1 1 0.8083 0.4
Circle phantom 0.004 Noise
SVMLS Method Proposed Method
Mean
error rate
Accurac
y
Precisio
n
Specificit
y
Sensitivit
y
Mean
error rate
Accurac
y
Precisio
n
Specificit
y
Sensitivi
ty Threshold
0.0276 0.9724 0.9334 0.9919 0.8302 0.023
1 0.9769 1 1 0.8080 0.2
Circle
0.004
noise
0.0309 0.9691 0.9047 0.9880 0.8305 0.023
5 0.9765 1 1 0.8048 0.3
0.0295 0.9705 0.9208 0.9903 0.8257 0.023
6 0.9764 1 1 0.8035 0.4
Circle phantom 0.008 Noise
SVMLS Method Proposed Method
Mean
error rate
Accurac
y
Precisio
n
Specificit
y
Sensitivit
y
Mean
error
rate
Accurac
y
Precisio
n
Specificit
y
Sensitivi
ty Threshold
0.0303 0.9697 0.9167 0.9898 0.8232 0.025
1 0.9749 1 1 0.7915 0.2
Circle
0.008 noise
0.0298 0.9702 0.9209 0.9903 0.8226 0.025
1 0.9749 1 1 0.7909 0.3
0.0284 0.9716 0.9249 0.9908 0.8311 0.023
9 0.9761 1 1 0.8013 0.4
S. Abdehzadeh et.al. / International Journal of Academic Research in Computer Engineering
64
each 8 slices of brain image selected and test has
been done on them and to convert to binary we use
thresholding, and for each image slices we have
done thresholding two time with values 0.3 and 0.4
and result of this thresholding were calculated for
SVMLS method, proposed method and manual
method(standard) and parameters calculated and
represented in Table 3 and 4 and 5.
Table 3. ROC Parameters Calculated By the Proposed Method and the Method SVMLS for Without Noise T1-MRI Image
T1-MRI Without Noise
SVMLS Method Proposed Method
Mean
error
rate
Accuracy Precision Specificity Sensitivity
Mean
error
rate
Accuracy Precision Specificity Sensitivity Threshold
0.0432 0.9568 0.9921 0.9929 0.9195 0.0156 0.9844 0.9817 0.9823 0.9866 0.3 T1-MRI Slice -
86 0.0467 0.9533 0.9926 0.9934 0.9117 0.0176 0.9824 0.9817 0.9823 0.9824 0.4
0.0423 0.9577 0.9920 0.9928 0.9213 0.0144 0.9856 0.9816 0.9821 0.9893 0.3 T1-MRI
Slice - 87
0.0443 0.9557 0.9922 0.9930 0.9170 0.0153 0.9847 0.9816 0.9821 0.9874 0.4
0.0437 0.9563 0.9918 0.9927 0.9186 0.0147 0.9853 0.9816 0.9821 0.9886 0.3 T1-MRI
Slice - 88
0.0458 0.9542 0.9922 0.9931 0.9140 0.0159 0.9841 0.9816 0.9821 0.9862 0.4
0.0438 0.9562 0.9917 0.9926 0.9185 0.0142 0.9858 0.9817 0.9822 0.9895 0.3 T1-MRI
Slice -
89 0.0458 0.9542 0.9923 0.9931 0.9139 0.0152 0.9848 0.9818 0.9823 0.9874 0.4
0.0464 0.9536 0.9924 0.9932 0.9125 0.0154 0.9846 0.9816 0.9821 0.9871 0.3 T1-MRI Slice -
90 0.0487 0.9513 0.9927 0.9936 0.9074 0.0164 0.9836 0.9816 0.9821 0.9851 0.4
0.0492 0.9508 0.9936 0.9944 0.9057 0.0134 0.9866 0.9815 0.9820 0.9913 0.3 T1-MRI Slice -
91 0.0513 0.9487 0.9940 0.9947 0.9011 0.0142 0.9858 0.9816 0.9820 0.9898 0.4
0.0569 0.9431 0.9944 0.9952 0.8892 0.0121 0.9879 0.9816 0.9820 0.9941 0.3 T1-MRI
Slice - 92
0.0601 0.9399 0.9947 0.9954 0.8823 0.0131 0.9869 0.9816 0.9820 0.9919 0.4
0.0605 0.9395 0.9960 0.9966 0.8804 0.0117 0.9883 0.9815 0.9819 0.9949 0.3 T1-MRI
Slice -
95 0.0626 0.9374 0.9963 0.9968 0.8758 0.0122 0.9878 0.9816 0.9820 0.9937 0.4
In Table 3 estimates the comparison and the results
of evaluation parameters at images without noise
for the proposed method and SVMLS method that
their results show that the classification accuracy of
the proposed method is better than SVMLS.
Table 4. ROC Parameters Calculated By the Proposed Method and the Method SVMLS for %3 Noise T1-MRI Image
T1-MRI 3% Noise
SVMLS Method Proposed Method Mean
error
rate
Accuracy Precision Specificity Sensitivi
ty
Mean
error rate
Accurac
y
Precisio
n Specificity Sensitivity
Thresho
ld
0.0405 0.9595 0.9917 0.9925 0.9254 0.0140 0.9860 0.9816 0.9821 0.9901 0.3 T1-MRI
Slice - 86 0.0430 0.9570 0.9922 0.9930 0.9198 0.0152 0.9848 0.9817 0.9822 0.9875 0.4
0.0430 0.9570 0.9910 0.9919 0.9208 0.0146 0.9854 0.9816 0.9821 0.9888 0.3 T1-MRI
Slice - 87 0.0455 0.9545 0.9917 0.9926 0.9150 0.0157 0.9843 0.9815 0.9821 0.9865 0.4
0.0430 0.9570 0.9907 0.9916 0.9212 0.0140 0.9860 0.9815 0.9820 0.9903 0.3 T1-MRI
Slice - 88 0.0452 0.9548 0.9915 0.9924 0.9159 0.0150 0.9850 0.9815 0.9820 0.9881 0.4
0.0434 0.9566 0.9915 0.9923 0.9196 0.0144 0.9856 0.9816 0.9821 0.9892 0.3 T1-MRI
Slice - 89 0.0458 0.9542 0.9921 0.9929 0.9141 0.0157 0.9843 0.9816 0.9821 0.9866 0.4
0.0472 0.9528 0.9919 0.9928 0.9114 0.0153 0.9847 0.9815 0.9820 0.9875 0.3 T1-MRI Slice - 90 0.0496 0.9504 0.9924 0.9933 0.9059 0.0165 0.9835 0.9815 0.9820 0.9849 0.4
0.0508 0.9492 0.9935 0.9943 0.9025 0.0138 0.9862 0.9815 0.9820 0.9905 0.3 T1-MRI
Slice - 91 0.0540 0.9460 0.9940 0.9947 0.8955 0.0151 0.9849 0.9815 0.9820 0.9879 0.4
0.0562 0.9438 0.9944 0.9951 0.8907 0.0135 0.9865 0.9815 0.9820 0.9911 0.3 T1-MRI
Slice - 92 0.0606 0.9394 0.9949 0.9956 0.8812 0.0150 0.9850 0.9815 0.9820 0.9880 0.4
0.0588 0.9412 0.9956 0.9962 0.8843 0.0111 0.9889 0.9816 0.9819 0.9961 0.3 T1-MRI
Slice - 95 0.0614 0.9386 0.9959 0.9965 0.8787 0.0116 0.9884 0.9816 0.9820 0.9950 0.4
S. Abdehzadeh et.al. / International Journal of Academic Research in Computer Engineering
65
In Table 4 estimates the comparison and the results
of evaluation parameters at images 3% noise for the
proposed method and SVMLS method the results
show that the classification accuracy of the
proposed method is better than SVMLS.
Table 5. Roc Parameters Calculate by the Proposed Method and the Method SVMLS for 9% Noise T1-Mri Image
T1-MRI %9 Noise
SVMLS Method Proposed Method Mean
error
rate
Accuracy Precision Specificity Sensitivity
Mean
error
rate
Accuracy Precision Specificity Sensitivity Thre
shold
0.0109 0.9891 0.9822 0.9959 0.9825 0.0103 0.9897 0.9816 0.9977 0.9819 0.3 T1-MRI Slice - 86 0.0452 0.9548 0.9911 0.9161 0.9921 0.0158 0.9842 0.9815 0.9865 0.9820 0.4
0.0107 0.9893 0.9827 0.9958 0.9830 0.0103 0.9897 0.9816 0.9977 0.9819 0.3 T1-MRI
Slice - 87 0.0471 0.9529 0.9911 0.9124 0.9921 0.0158 0.9842 0.9815 0.9864 0.9820 0.4
0.0144 0.9856 0.9834 0.9874 0.9839 0.0105 0.9895 0.9816 0.9974 0.9819 0.3 T1-MRI
Slice - 88 0.0443 0.9557 0.9911 0.9181 0.9920 0.0141 0.9859 0.9816 0.9898 0.9820 0.4
0.0206 0.9794 0.9861 0.9718 0.9867 0.0107 0.9893 0.9816 0.9968 0.9819 0.3 T1-MRI
Slice - 89 0.0468 0.9532 0.9912 0.9129 0.9922 0.0159 0.9841 0.9816 0.9862 0.9821 0.4
0.0136 0.9864 0.9839 0.9886 0.9843 0.0104 0.9896 0.9816 0.9976 0.9819 0.3 T1-MRI
Slice - 90 0.0500 0.9500 0.9926 0.9050 0.9934 0.0154 0.9846 0.9815 0.9874 0.9820 0.4
0.0136 0.9864 0.9844 0.9879 0.9848 0.0104 0.9896 0.9816 0.9976 0.9819 0.3 T1-MRI Slice - 91 0.0544 0.9456 0.9929 0.8957 0.9938 0.0149 0.9851 0.9815 0.9883 0.9820 0.4
0.0126 0.9874 0.9842 0.9903 0.9846 0.0103 0.9897 0.9816 0.9977 0.9819 0.3 T1-MRI Slice - 92 0.0587 0.9413 0.9942 0.8856 0.9950 0.0137 0.9863 0.9815 0.9909 0.9819 0.4
0.0356 0.9644 0.9915 0.9356 0.9922 0.0104 0.9896 0.9816 0.9975 0.9819 0.3 T1-MRI
Slice - 95 0.0615 0.9385 0.9954 0.8788 0.9960 0.0119 0.9881 0.9817 0.9942 0.9821 0.4
In Table 5, estimates the comparison and the results
of evaluation parameters at images 9% noise for the
proposed method and SVMLS method the results
show that the classification accuracy of the
proposed method is better than SVMLS. At the
following, we will compare and study the
important performance of ROC curve parameters
that the results and the obtained data are
represented at 6, 7 and 8.
Table 6. Performance Measurement Parameters By
Proposed Method And The Method SVMLS For Without
Noise T1-MRI Image
T1-MRI Without Noise
Propose
d
Method
(%)
Svmls
Metho
d
)%(
Definition Measure
98.90 90.55 TP/(TP+FN) Sensitivit
y
98.21 99.39 TN/(TN+FP) Specificit
y
98.16 99.31 TP/(TP+FP) Precision
98.55 95.05 (TP+TN)/(TP+TN+FP+FN
) Accuracy
1.44 4.94 (FP+FN)/(TP+TN+FP+FN
)
Mean
error rate
According to Table 6, it is determined that the performance measurement of the proposed method was better in terms of classification accuracy. Also to better assess and view the results, we can see the results on graphs in Figure 13.
Table 7. Performance measurement parameters by
proposed method and the method SVMLS for %3 noise
T1-MRI image
T1-MRI 3% Noise
Proposed
Method
(%)
Svmls
Method
)%(
Definition Measure
98.92 90.63 TP/(TP+FN) Sensitivity
98.20 99.36 TN/(TN+FP) Specificity
98.15 99.28 TP/(TP+FP) Precision
98.55 95.07 (TP+TN)/(TP+TN+FP
+FN) Accuracy
1.44 4.92 (FP+FN)/(TP+TN+FP
+FN)
Mean
error rate
According to Table 7, it is determined that the performance measurement of the proposed method was better in terms of classification accuracy. Also to better assess and view the results, we can see the results on graphs in Figure 14.
Table 8. Performance measurement parameters by
proposed method and the method SVMLS for 9% Noise
T1-MRI image
T1-MRI 9% Noise
Proposed
Method
(%)
)%(
Svmls
Method
Definition Measure
98.19 98.92 TP/(TP+FN) Sensitivity
99.31 94.23 TN/(TN+FP) Specificity
98.15 98.86 TP/(TP+FP) Precision
98.74 96.62 (TP+TN)/(TP+TN+FP+FN) Accuracy
1.25 3.37 (FP+FN)/(TP+TN+FP+FN) Mean
error rate
S. Abdehzadeh et.al. / International Journal of Academic Research in Computer Engineering
66
According to Table 8, it is determined that the
performance measurement of the proposed method
was better in terms of classification accuracy. Also
to better assess and view the results, we can see the
results on graphs in Figure 15. ROC curves for
eight noise-free image of the brain with the
proposed method showed in figure 13 and with 3%
and 9% noise showed in Figure 14 and 15
respectively.
Figure 13. ROC curves for eight without noise image of
the brain with the proposed method and SVMLS
Figure 14. ROC curves for eight with 3% noise image of
the brain with the proposed method and SVMLS
Figure 15. ROC curves for eight with 9% noise image of
the brain with the proposed method and SVMLS
5. Conclusions and Future Works In this paper, a multi-stage level set method and the active contour was presented so that the bias correction and images segmentation is able to be done simultaneously. Smoothing bias is inherently guaranteed without any additional cost by using normalized convolution. According to the generally accepted model, images with non-uniform intensity and obtained local intensity clustering features, we define an energy function of the level set functions that shows one part of image domain and a bias field for none-uniform intensities. Thus, the segmentation and bias field estimation by minimizing an energy function to be implemented jointly. Slowly changing bias field feature obtained from the proposed energy naturally given expression is guaranteed in our framework, Without having to enter a clear Flattening value in bias field. Finally make the method stronger by combining with active contour model in order to obtain more accurate results. The results showed that this method improves the accuracy more than the other methods.
References
[1] M., Zeydabadi M., aghayezade R. and
soltanyan H. Segmenting multi-resolution
magnetic resonance images of the human
brain. The fifth Conference Intelligent
Systems, 2003.
[2] M., Hosntalab , F.B., Mofrad, N., Mnshizadeh
and M., Amoui, Automatic Left Ventricle
Segmentation in Volumetric SPECT Data Set
By Vibrational Level Set, International Journal
of Computer Assisted Radiology and Surgery,
Vol. 7, No,6, pp.837-843, 2012.
[3] H.,Aslian M., Sadeghi, S.R., Mahdavi F.B.,
Mofrad, M.,Astarakee, N., Khaledi and P.
Fadavi, Magnetic Resonance Imaging-Based
Target Volume Delineation in Radiation
Therapy Treatment Planning for Brain Tumors
Using Localized Region-Based Active
Contour, International Journal of Radiation
Oncology Biology Physics, Vol. 87, Issue 1,
pp. 195-201, 2013.
[4] M., Sezgin, B., Sankur, Survey over image
thresholding techniques and quantitative
performance evaluation. Journal of Electronic
Imaging, Vol. 13, No.1, pp. 46-165, 2004.
[5] F.S. Gharehchopogh, S. Ebrahimi, A Novel
Approach for Edge Detection in Images based
on Cellular Learning Automata, International
Journal of Computer Vision and Image
Processing (IJCVIP), Vol. 2, No. 4, pp. 53-64,
October-December 2012.
S. Abdehzadeh et.al. / International Journal of Academic Research in Computer Engineering
67
[6] T.Y., Kuo, Y.Y., Lai and Y.C., Lo, A Novel
Image Binarization Method Using Hybrid
Thresholding. International Conference on
Multimedia and Expo (ICME), IEEE, pp. 608-
612, 2010.
[7] M., Kass, A.,Witkin and D.,Terzopoulos,
Snakes: active contour models. International
Journal of Computer Vision, Vol.1, No.4,
pp.321–331, 1987.
[8] C.,Rigaud, J.C., Burie, J.M., Ogier, D.,
Karatzas, and J.V.D. Weijer, An Active
Contour Model for Speech Balloon Detection
in Comics. 12th International Conference on
Document Analysis and Recognition, IEEE,
pp. 1240-1244, 2013.
[9] L.D., Cohen, On active contour models and
balloons. CVGIP: Image Understanding,
Vol.53, No. 2, pp. 211-218. 1991.
[10] A., Dervieux, F., Thomasset F, A Finite
Element Method for The Simulation of
Ayleigh-Taylor Instability, Lecture Notes in
Mathematics, Springer Berlin Heidelberg,
pp.145-158, 1980.
[11] S., Osher and J. Sethian, Fronts Propagation
with Curvature Dependent Speed: Algorithms
Based on Hamilton-Jacobi Formulations.
Journal of Computational Physics, Vol.75, No.
1, pp.12-49, 1988.
[12] R.,Tsai and S.,Osher, Level Set Methods and
Their Applications in Image Science.
Communications in Mathematical Sciences,
Vol. 1, No.3, pp.1–20, 2003.
[13] L.A., Vese and T.F., Chan, A Multiphase Level
Set Framework for Image Segmentation Using
the Mumford and Shah Model. International
Journal of Computer Vision, Vol.50, No.3, pp.
271-293, 2002.
[14] K., Zhang, L., Zhang and S. Zhang, A
Vibrational Multiphase Level Set Approach to
Simultaneous Segmentation and Bias
Correction, International Conference on Image
Processing, IEEE, pp.4105 – 4108, 2010.
[15] Dataset., Available:
http://brainweb.bic.mni.mcgill.ca/brainweb/
[Last Accessed 7 May 2016].
[16] B., Widrow, The Rubber-Mask Technique,
Pattern Recognition, pp.175– 211, 1973.
[17] C. Li, R., Huang, Z., Ding, C., Gatenby, D.,
Metaxas and J. Gore, A Vibrational Level Set
Approach to Segmentation and Bias Correction
of Medical Images with Intensity
Inhomogeneity, Medical Image Computing
and Computer-Assisted Intervention–
MICCAI, Springer Berlin Heidelberg, 1083–
1091, 2008.
[18] Y., Yuan and C., He, Adaptive active contours
without edges. Mathematical and Computer
Modelling, Vol.55, No.5, pp. 1705-1721, 2012.
[19] T.F., Chan, B.Y., Sandberg and L.A. Vese,
Active Contours without Edges for Vector-
Valued Images. Journal of Visual
Communication and Image Representation,
Vol. 11, No.2, pp.130–141, 2000.
[20] F., Tony Chan and A.L., Vese, Active Contours
Without Edges. IEEE Transactions on Image
Processing, Vol. 10, No.2,pp. 266-277, 2001.