Segmentation of Brain MRI Using Active Contour Model

Amira Ben Rabeh, Faouzi Benzarti, Hamid Amiri

National Engineering School of Tunis (ENIT), Signal, Image and Technology Information (LR-SITI),
Manar, Tunisia

Received 26 April 2016; revised 28 October 2016; accepted 5 December 2016

ABSTRACT: Alzheimer disease is a neurodegenerative disorder that the most affected, but the dementia is striking more and more young
impairs memory, cognitive function, and gradually leads to dementia, people. What constitutes a prime therapeutic challenge is the fact
physical deterioration, loss of independence, and death of the affected that no current treatment can permanently eradicate the disease.
individual. In this context, segmentation of medical images is a very (Aminoff et al, 2007; Bar et al, 2003; Bear et al, 2007; Bollinger et
important technique in the field of image analysis and Computer-
al, 2011; Buckner et al , 2008; Dickerson et al, 2009; Epstein et al,
Assisted Diagnosis. In this article, we introduce a new automatic method
of brain images segmentation based on the Active Contour (AC) model
1999). In fact, it is very important to go through an image processing
to extract the Hippocampus and the Corpus Callosum (CC). Our contri- to detect and locate the affected area. The image segmentation is one
bution is to combine the geometric method with the statistical method of of the techniques that can bring about some crucial information for
the AC. We used the Caselle Level Set and added a learning phase diagnostic assistance. It is defined as the process of partitioning a
to build an average shape and to make the initialization task automatic. digital image into multiple segments or regions.
For the step of contour evolution, we used the principle of Level set The diversity of images, the difficulty of the problems, the
and we added to it the a priori knowledge. Experimental results are researchers varied origins, the evolution of computers, as well as
very promising. V C 2017 Wiley Periodicals, Inc. Int J Imaging Syst Technol, 27,
certain empiricism in the evaluation of results led to the introduction
311, 2017; Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: of a multitude of algorithms. The reliable and accurate segmentation
10.1002/ima.22205
of anatomical volumes whether normal or pathological from imaging
Key words: Alzheimer disease; computer-assisted diagnosis; hippo- systems aims at developing the medical diagnosis because it may
campus; corpus callosum; active contour help studying morphology internal structures, and it may help the
detection and quantification of the lesions. Many segmentation meth-
ods have been proposed for many years for the MRI type of images.
I. INTRODUCTION The automatic brain MRI segmentation is a complex problem from
Alzheimers disease is a neurodegenerative disorder that impairs the point of view of the variability of the human brain in terms of
memory and cognitive function. It is divided into three main phases: images processing acquired by this method. Brain MRI segmentation
the Mild Cognitive Impairment (MCI) stage, the moderate stage, and is an essential basic step that has many applications in neurology
the severe stage. At the mild stage, the person is usually able to think such as quantitative analysis, operational planning, and functional
appropriately, and to make crucial decisions about future; However, imaging. Although very accurate to describe the structures of the
Some negative symptoms are remarkable during this stage such as brain, magnetic resonance images present many difficulties like low
forgetfulness and communication problems, for instance, finding the contrast and inhomogeneity, but their treatment can be achieved
right word or following a conversation. In the moderate stage, there through simple approaches. In this work, we suggest a new automatic
is a further deterioration of mental and physical abilities of the per- method of segmentation based on the Active Contour (AC) model.
son. Memory loss worsens. One can no longer remember his/her per- This article is organized as follows. In the following section, we
sonal history and can no more recognize his/her family and friends. present some related works. In the third section, we present the sug-
In the severe stage, the person is no longer able to remember any- gested method. In the last section, we give some experimental
thing, communicate with anyone, or even take care of oneself. The results, and a discussion.
patient will require a special care day and night. This disease can be
genetic, but that is not the primary cause of degenerative dementia. II. RELATED WORKS
The latest statistics show that there are 12 million patients world-
There are many methods for segmentation of the Brain MRI,
wide. This number keeps growing and the forecast estimates 18 mil-
such as:
lion cases in 2025. In Tunisia, the number is alarming: about 30 000
people are affected, to 1.5 million in the Arab world. The elderly are
A. ASM 1 D Proposed by (Said Taieb et al., 2014). This
Correspondence to: Amira Ben Rabeh; e-mail: [email protected] method proposed by Said Taieb et al. (2014) to extract the

C 2017 Wiley Periodicals, Inc.

V
Hippocampus using the static model: ASM (Active Shape Model) this A.1. Average Shape and Automatic Position. Preparation of
method of segmentation is based in the priori knowledge. He used the the learning base The preparation of a training base for a given
distance between the two hippocampus as a priori knowledge. structure is to initially collect a sample image database reflecting
possible variations in its shape. For that, the constructed model
B. Method Proposed by (Kayalvizhi et al., 2013). This would be sufficiently representative, the number of images collected
method proposed by Kayalvizhi et al. (2013), he analyzed MRI brain should be relatively high and instances of quite varied shapes should
images: coronal section weighted T1 and studied the progression of be selected. Then, we extract the shapes from the sample images on
severity in Alzheimers disease conditions. He used the geometric each image by placing a sufficient number of characteristic points.
model Level Set. These points should be placed carefully on the contour of the struc-
ture studied to clearly represent its details. The points, which are
C. Method Proposed by S. Matoug. As a solution, he defined by an expert, provide a precise meaning. In fact, each form
introduced an automatic tool in 2011 that reads volumetric MRI and will be modeled by a vector 3 constructed by concatenation of the
performs two-dimensional volume slices and volumetric segmenta- coordinates of characteristic points located on the outline:
tion methods to segment gray matter, white matter, and cerebrospinal
fluid. For segmentation, he used the parametric model Snake. x5x1; x2; . . . . . . . . . :; xn (1)

The training set can be modeled by a set of vectors:

III. PROPOSED METHOD
We used the MRI data sets database from the Open Access Series of fxig avec i51 . . . N N number of sample images
Imaging Studies (Ardekani and Bachman, 2009).
Our contribution is based on: Forms alignment The extracted forms during the previous step
are independent of the parameters of size, and position. There may
a. Preparation of a learning phase to extract an average shape
have shapes, which are more or less remote in different directions
and automatic position of this form.
and have different sizes. However, the modeling approach is to study
b. Improvement at the stage of the deformable model evolution.
only the essential shape variation between the different configura-
We incorporated a priori knowledge of the surface and the
tions of the studied structure. Thus, to solve this problem of variation
standard deviation.
of size and position of the N forms of the training set, the contours
Overall model: we will integrate a priori knowledge of surface must be aligned with one of the obtained contours. This will be used
and deviation of the studied area (Fig. 1). as a reference.
The alignment procedure is to first take a random form which
A. Learning Phase. The objective is to learn the geometry and is aligned with all others. Then, at each iteration, an average shape
the variations in the structure that it is desired to locate later. There- is calculated, and normalized on the other which will be realigned.
fore, it seeks to establish a model that describes the shape and typical This process is stopped when the average form reaches a certain
fluctuations. This first step requires the preparation of a learning criterion. It is supposed to converge regardless of the initial form
database reflecting possible variations in shape of this structure. chosen.

Figure 1. Overall model. [Color figure can be viewed at wileyonlinelibrary.com]

4 Vol. 27, 311 (2017)

Generation of the average shape The aligned vectors, resulting Vs5 S1; S2; . . . ; Sn
from the previous two steps can be arranged as a matrix of size (2n,
N), called the observation matrix. The columns correspond to shapes The average surface of interest structures
and lines correspond to the coordinates of the characteristic points
describing each shape. 1X n

This can be done by examining the variation of homologous char- Sm5 Si (6)
n i51
acteristic points on all forms. If we consider, for example, the first
two lines of the observation matrix, they represent the coordinates of
the first feature point through different forms. By examining the
change in these coordinates and passing from one form to another,  The variance, which measures the dispersion of the elementary
one can deduce the path followed by this point and the amplitude of surfaces (Si) relative to the average surface:
variation. However, the problem is that each form is represented by
2
n feature points, each identified by its coordinates (x, y) in the plan. 1X n
Vs5 Si2Sm (7)
Therefore, for a single point, we used 2N variables and for all the n i51
points, we used 2nN variables. To simplify the problem and reduce
the number of variables that describe each shape, the idea is to apply
a principal component analysis (PCA) on the raw data.  The standard deviation, which represents the average of all devi-
In practice, the main components are derived from the covariance ations; basic variances compared with the average distance:
matrix associated with the observation matrix. To determine the p
mode of variation of a shape, we can simply compare this shape with r5 Vs (8)
respect to an average computed shape, instead of comparing with all
other forms. This comparison is between the coordinates of points as
Then, we will propose a confidence interval [Sm 2 2r, Sm 1 2r].
characteristics compared with counterparts in average shape. The
For each iteration, we calculate the surface of the area. If the
deviation of the ith shape with respect to the average shape may be
surface is lower than the lower limit of the interval, we stop the
defined by the expression:
evolution of the curve.
dxi5xi1 ; yi1 ; xi2 ; yi2 ; . . . ; xin ; yin 2x1 ; y1 ; x2 ; y2 ; . . . ; xn ; yn 5xi2x A.3. Priori Knowledge of Standard Deviation. Each image is
(2) recorded by the variation of the studied area.
We integrate the prior variance. We search the average, the
XN
When x5 N1 xi, the average computed shape from the learning variance, and the standard deviation of the variance of form.
i51
base aligned.
Vars 5 Var1; Var2; . . . ; Varn (9)
Automatic initialization of the average shape Our first contribu-
tion is to overcome the stage of manual initialization. In this way: Constructing a vector contains all the variations of the areas of the
0 1
X1Y1 0 1 hippocampus of the samples of the base.
B C p1 The average change in interest structures:
B X2Y2 C B
B C B p2 C C
B C
B.........C B C
1X n
B C B C
B C B
B . . . :: C
C
Varm5
n i51
Vari (10)
M5B C
B . . . . . . . . . C5B C (3)
B C B
B . . . :: C
C
B.........C B C
B C B The variance, which measures the dispersion of elementary varia-
B C @ . . . :: C
A
B XNYN C tions (Si) relative to the average variation.
@ A
pN
1X n
Vvar5 Vari2Varm2 (11)
N 5 n*Nech n i51
n: number of points that have the outline of an image.
Nech: number of samples The greater the variance is, the more heterogeneous the appropriate
The M matrix is of dimension N * 2 area (hippocampus) is. The standard deviation, which represents the
Through this matrix, the average position is determined by: average of all deviations basic variances compared with the average
distance.
1X N
Posm5 pi (4) p
N i51
r5 Vvar (12)
A.2. Priori Knowledge of the Surface. Each image is recorded Then, we will propose a confidence interval [Varm 2 2r, Varm 1
by the surface of the studied form by using Eq. (5): 2r].
The goal is to derive a compact formulation describing the per-
X
nbl X
nbc
Si5 Xij (5) mitted variance. In fact, from Vars, we can calculate the parametric
i51 j51 statistics. The basic idea is if there is a level change out of range, it
offers the inverse of the curve. The curve evolves internally instead
Xij: the pixel is in the iem row and column jem. of the convergence towards the outer neighbor pixels.

Vol. 27, 311 (2017) 5

B. Segmentation. B.1. Development of Method. Statistical In addition to that, they allow to detect several contours at a time
method (based on priori knowledge) and geometric method (The and this takes the same amount of time as for the calculations of the
evolution of the curve based on neighbored pixels). In the following 1 level function. However, despite their advantages, these models
sections, we will first introduce each method apart and then, we will have limitations. First of all, they are only valid for closed curves,
explain why and how we combined the two of them (Aljabar et al., which greatly limits their use. Then the resolution procedure is rather
2009; Atif et al., 2006; Ardekani et al., 2009; Delingette et al., 2001; difficult to implement, compared with parametric models. (Caselles
Duvernoy et al., 2005). et al., 1997; Chan and Vese, 2001; Charmi et al., 2008; Chen and
B.2. Geometric Deformable Models. The geometric models pro- Radke, 2009).
vide an implicit representation of the curve as the zero level of a sca- B.3. Statistical Deformable Models. The ASM is a variation of
lar function of higher dimension. The move towards the target statistic deformable models, introduced by Cootes and Taylor (1994)
contour is based on the evolution of curves theory, which only to extract objects complex and not rigid. The advantage of the ASM
depends on the intrinsic geometric properties. These properties are compared to other variants of deformable models lies in the fact that
the standards defining the direction of evolution and the curve defin- the evolution of the curve is guided by a strong a priori knowledge
ing the velocity at each point. Thus, no functional energy is used, about the geometry and structure of the deformation modes studied.
and therefore no programming is considered. This knowledge is represented by a statistical model, which form
Level set The Level Set method was originally introduced by describes a space of allowed deformations. The construction of such
Osher and Sethian. This method allows changing a curve implicitly a model is made by applying a PCA on a learning base, which
defined as the zero level of a higher order function to the target includes the various possible forms of the object.
curves in the image. Changes can be made according to several crite- These models provide a discrete representation of the curve by
ria. The formulation of this method, where the change is based on a set of feature points. They are based on a preliminary statistical
the average intensities inside and outside of the curve, is well suited analysis of the variations of the designed structure. The implementa-
to the detection of homogeneous regions having no sharp boundaries. tion of a statistical model then requires the preparation of N represen-
It is also more robust in the presence of noise. tative examples of the object under study and the priori knowledge.
The level set model (Level set) suggests an interesting analytical The shape of the object in each case is modeled by a vector of feature
framework for geometric curve evolution through the implicit repre- points x. Then, an average shape x supposed to be a relevant repre-
sentation. It describes the curve as the zero level of a scalar function sentative of all forms is deduced:
of / higher dimension levels, instead of using a conventional para-
metric representation. The function of / levels will be defined in the 1X N
image plan and will evolve over time in the direction of normality x5 xi (16)
N i51
and with a speed of propagation. Therefore, the spread (zero curve)
will be detected easily from that of /. The latter is defined as a
By applying a PCA on the average shape x. Patterns and magnitudes
signed distance between any point of the image and the curve:
of deformation of the object studied are detected. Thus, a particular
form x can be simply expressed in terms of the average shape and a
/ : R2 ! R
linear combination of the principal modes of deformation:
/x; y; t57dx; y; v (13)
x5x1/b (17)
/x; y; t50 : six; y 2 v Distance fromx; y to v
When /5/1 ; /2 ; /3 ; . . . /t is an orthogonal basis set by the modes
The initial curve V0 is defined by points such as: /x; y; t5050
of t the most significant deformations (principal components).
The evolution equation of the curve Vthrough the function / is
b5b1 ; b2 ; b3 ; . . . bt t Represents the projection of a form x in
defined by:
this base. By varying the parameters, new forms similar to those
@/v; t obtained in the learning phase can be generated. Therefore, an area
1Fk jr/v; tj50/v; t50 (14) of authorized forms can be defined by properly limiting parameters:
@t

During evolution, all forms represented by the curves can be defined 2bmax i  bi  bmax i ; i5 1 . . . t (18)
by:
The final outline, describing the shape of the target object, must
fx; y=/x; y; t50g (15) belong to this space. In the localization process, the evolution of the
initial curve (which is usually the average shape) is done through a
With this implicit representation, this model provides an opportunity strategy using local intensity of the image. This development can be
to change the topology, and thus the detection of multiple contours done in the space of predefined authorized deformations.
from a single initial curve. However, the numerical implementation The ASM tenders ability to integrate a strong knowledge very
of the 1 level function is very time-consuming. Indeed, the evalua- effectively prior shape on the structures of interest. It is based on sta-
tion of 1 involves estimating local features (normal curvature) at tistical models of shape structures designed to guide the segmenta-
any point of the image. Several other similar models, using an tion step. This reduces solutions space and always lead to correct
implicit representation of the curve and adopting geometric con- forms (Andreopoulos et al., 2008; Aymeric et al., 2005; Babalola et
straints for evolution have been proposed. This implicit representa- al., 2009; Bascle et al., 1994).
tion explains all the success of this class of geometric models. B.4. Fusion of the Two Models. In the learning step, we used
Indeed, the theoretical results have shown the existence and the the principle of the Statistical model; we used n patterns and we
uniqueness of the solution, which depends on few parameters. determined the average shape and the average position. In addition,

6 Vol. 27, 311 (2017)

 Figure 2. Improvement using variation. [Color figure can be viewed at wileyonlinelibrary.com]

Figure 3. Improvement using surface. [Color figure can be viewed at wileyonlinelibrary.com]

we calculated the two confidence intervals to integrate the priori Body We launch the initial curve.
knowledge. For each iteration, calculating the surface and the variation:
For the evolution of the curve, we used the principle of the geo- If the surface is higher than the upper limit of the interval:
metric model: evolution of the curve based on neighbored pixels.
 Inverse the changing curvature. (To ensure that the outline
We combined the two models, the geometric and the statistical to
does not converge to other areas that appear to be consistent
integrate the priori knowledge to the geometric model, and therefore,
the segmentation method will be automatic. The objective of the two with the area studied) and stop the evolution of the curve.
models fusion is to improve the segmentation quality. The curve can If the surface is in the range of trust:
converge towards other homogeneous areas with the studied area.
Integrating the priori knowledge of the surface, we can improve the  The daily evolution is kept.
limit. The curve can converge towards other inhomogeneous areas If the variation exceeds the upper range of the confidence
with the studied area via integrating the priori knowledge of the vari- interval:
ation. We can improve the limit of the segmentation.
Basic algorithm Once the learning phase is done, we get two con-  A reverse development is provided, the curve changes internal-
fidence intervals: For each iteration, we extract the area studied (Hip- ly instead of converging toward the outer neighbor pixels.
pocampus or the Corpus Callosum, CC) from contour. Then, we Output We find the final curve and we extract the correct area.
calculate the surface. If the surface is included in the interval, we fol-
low the daily evolution, otherwise if the surface exceeds the upper
IV. EXPERIMENTAL RESULTS
bound, the evolution stops. We calculate the variation of the studied
area. If the variation exceeds the upper limit, the daily changes are We are interested to extract the two areas: Hippocampus and Corpus
reversed, so the curve evolves internally instead of converging Callusm. We used three subjects: normal, suffering from Alz-
toward the outer neighbor pixels. heimers disease in the early stage and another in an advanced stage.
Using the constraint of the variation to improve the limit by the We used the two sections frontal and sagittal for any subjects. We
figures below: the contour may include the hippocampus more areas segmented the hippocampus of the three subjects by four methods:
surrounding it and that are not homogeneous with the desired area. parametric model (snake; Charmi et al., 2008), static model (ASM;
Through the confidence interval of variation and a priori knowledge, Cootes and Taylor, 1994), geometric model (Level Set; Caselles
these limitations can be overcome (Fig. 2). et al., 1997) and the fourth method is our proposed method Modify
The constraint of the surface is used to improve the limit by the Level set which is a combination between the static and geometric
figures below: the contour may include the hippocampus more areas models. We show the results in the figure below (Fig. 4).
surrounding it and that are homogeneous with the desired area. Using our method Modify Level Set, we can calculate some fea-
Through the surface of confidence interval and a priori knowledge, tures of geometries such as surface, perimeter, and standard devia-
these limitations can be overcome (Fig. 3). tion. We present in the table below the measurement of the features
The algorithm in below presents the principle of our method: for the two parts of hippocampus (Tables I and II).
Input After the step of the training:
A.1. EXTRACTION OF THE CC
 We determine the two confidence intervals of the Surface and We segmented the CC of the three subjects by four methods (Fig. 5).
the variation. In reviewing the results: for lines 3, 4, and 5, we note that the
 We determine the average shape and the initial curve. curve migrates to similar intensity regions and sometimes not. For

Vol. 27, 311 (2017) 7

Figure 4. Results of the segmentation of the hippocampus. The six lines present in order: initial image, image manually segmented, the result
using the snake method (parametric model), the result using the ASM method (statistical model), the result using the method Level Set
(geometric model), the result using the Modify Level Set method (geometric model 1 Statistics). Column 1 shows a normal subject, column 2
about Alzheimer (primary stage) and the third column corresponds to an Alzheimers subject (advanced stage). [Color figure can be viewed at
wileyonlinelibrary.com]

Table I. Table of measurement for the hippocampus.

Hippocampus L Hippocampus R
Surface Perimeter Standard Deviation Surface Perimeter Standard Deviation
Subject 1 1039 174.61 7 1425 170 8.3
Subject 2 1019 166 14 1267 160 17
Subject 3 902 155 17 1111 151 18.7

8 Vol. 27, 311 (2017)

Table II. Table of measurement for Corpus Callosum. the 6th line, we can see that the curve evolves maintaining more
regular results. It happens to capture the shape of the target
Surface Perimeter Standard deviation
structure.
Subject 1 5518 687 7.4 In our case, we have three topics, segmented manually by a
Subject 2 5173 658 17 nuclear medicine department radiologist Kassab Hospital of Tunis.
Subject 3 3435 645 19
We will use this data to make a quantitative assessment of the pro-
posed model.
We integrated the a priori knowledge with the geometric mod-
el. We decided to compare our model Modify Level Set, the

Figure 5. Results segmentation corpus callosum. The six lines present in order: initial image, image manually segmented, the result using the
snake method (parametric model), the result using the ASM method (statistical model), the result using the method Level Set (geometric model),
the result using the Modify Level Set method (geometric model 1 Statistics). Column 1 shows a normal subject, column 2 about Alzheimer (primary
stage) and the third column corresponds to an Alzheimers subject (advanced stage). [Color figure can be viewed at wileyonlinelibrary.com]

Vol. 27, 311 (2017) 9

 The poor performance of the ASM model and the SNAKE model
can be explained by the vagueness of the initialization in some pic-
tures of the sequence and the generality of the prior information of
the used form.
Another interesting finding is that the value of the Hausdorff dis-
tance, such as Level Set, Modify Level set, or SNAKE compared
with manual segmentation, varies considerably from one sequence to
Figure 6. The figure shows the results of calculating the Hausdorff another. For example, in the Figure 6 this distance exceeds 23 mm
distance between the four methods (Modify Level Set, Level Set, for ASM and the SNAKE while in Figure 8 this distance does not
ASM, and Snake) compared with the ground truth for the normal sub-
exceed 16 mm. This may be related to the quality of the processed
ject based on iterations. [Color figure can be viewed at wileyonlineli-
brary.com]
sequence, which influences the result.
Through these measures, although in some cases the ASM model,
the model SNAKE provide more or less acceptable results, we can
see that our model for each sequence provides a more stable and
comprehensive income closer to the manual segmentation. This can
reveal the importance of the integration of a priori knowledge with
the geometric model.
In this work, we presented a segmentation method based on Level
Set formalism coupled with a priori knowledge of the form. This
work is part of the same knowledge integration problem a priori to
Figure 7. The figure shows the results of calculating the Hausdorff
improve segmentation results.
distance between the four methods (Modify Level Set, Level Set, Indeed, we have introduced two points value in the segmentation
ASM, and Snake) compared with the ground truth for the MCI subject process:
based on iterations. [Color figure can be viewed at wileyonlinelibrary.
 The initial zero level curve is calculated from the average
com]
shape of the structure to be segmented. It will be more close to
the target structure when locating.
 The change is limited to an allowed deformation field prede-
fined. This reduces space solutions and always result in eligible
forms.
This is what has led to promising segmentation method, which
was applied for segmentation of internal brain structures in MRI
images. That shows that the contributions of the integrated form of
statistical constraint improve the accuracy of results.
Figure 8. The figure shows the results of calculating the Hausdorff In conclusion, it is clear that the integration of a priori knowl-
distance between the four methods (Modify Level Set, Level Set, edge of the area and the changes of the shape of the hippocampus
ASM, and Snake) compared to the ground truth for the Alzheimer had significantly improved results for the segmentation of image
subject based on iterations. [Color figure can be viewed at wileyonli- sequences. This is what inevitably increases the reliability of
nelibrary.com] diagnostic parameters, which is calculated based on these
results.
However, it must be said that these findings and results can be
original geometric model Level Set, statistical model ASM and more inclusive and contain more sequences in the quantitative vali-
the method based on parametric AC (Snake) with the ground dation process. It should, also, be noted that the most delicate phase
truth. in the use of our approach is the form of modeling phase. This phase,
The four methods are based on the deformable contour; we chose which is based on a manual process, influences largely on the quality
the Hausdorff distance as a measure of quality of segmentation. This of results. It must, therefore, be carried out carefully with the help of
metric is widespread in the medical field and admits multiple appli- a domain expert.
cations. In our case, we will use this distance to measure the degree
of similarity between two geometric shapes (Figs. 68).
VI. CONCLUSION
In this article, we proposed a new automatic method of brain images
segmentation based on the AC model to extract the Hippocampus
V. DISCUSSION and the CC. Our contribution is to combine the geometric and statis-
Examining the three diagrams, we can clearly see that the green tical method of the AC. The evolution of the contour is based on the
curve shows some stability compared with the other curves; the blue, geometry and some priori knowledge such as surface and standard
the gray and the red. In fact, for the green line in the three diagrams, deviation. We find that our method Modify Level set give a good
the values of the Hausdorff distance are between 1,41et 4.12 (mm).In result. We can improve our work by adding a classification step. We
contrast, for the other three curves, the values of the Hausdorff dis- can use the method of classification Support Vector Machine. With
tance often show large variations, which rose from 4.26 to the two areas Hippocampus and CC, we can diagnose Alzheimer
38.72 mm. diseases.

10 Vol. 27, 311 (2017)

REFERENCES V. Caselles, R. Kimmel, and G. Sapiro, Geodesic Active Contours, Int J
P. Aljabar, R.A. Heckemann, A. Hammers, J.V. Hajnal, and D. Rueckert, Comput Vis 22 (1997), 6179.
Multi-atlas based segmentation of brain images: Atlas selection and its T. Chan and L. Vese, Active contours without edges, IEEE Trans Image
effect on accuracy, Neuroimage 46 (2009), 726738. Process 10 (2001), 266277.
E. Aminoff, N. Gronau, and M. Bar, The parahippocampal cortex mediates M.A. Charmi, S. Derrode, and F. Ghorbel, Fourier-based geometric shape
spatial and nonspatial associations, Cereb Cortex 17 (2007), 14931503. prior for snakes, Pattern Recognit Lett 29 (2008), 897904.
A. Andreopoulos and J. Tsotsos, Efficient and generalizable statistical mod- C.H. Chen and G.G. Lee, Image Segmentation Using Multiresolution Wave-
els of shape and appearance for analysis of cardiac MRI, Med Image Anal let Analysis and Expectation-Maximization (EM) Algorithm for Digital
(2008), 335357. Mammography, Int J Imaging Syst Technol 8 (1997), 491504.
B.A. Ardekani and A.H. Bachman, Model-based automatic detection of the W. Chen and M.L. Giger. A fuzzy c-means (FCM) based algorithm for
anterior and posterior commissures on MRI scans, Neuroimage 46 (2009), intensity inhomogeneity correction and segmentation of MR images. IEEE
677682. International Symposium on Biomedical Imaging, Arlington, VA, 2004, pp.
J. Atif, O. Nempont, O. Colliot, E. Angelini, and I. Bloch. Level Set 1307-1310.
Deformable Models Constrained by Fuzzy Spatial Relations. In Information S. Chen and R.J. Radke. Level set segmentation with both shape and intensi-
Processing and Management of Uncertainty in Knowledge-Based Systems, ty priors. 12th IEEE International Conference on Computer Vision, Japan,
IPMU, France, 2006, pp. 15341541. 2009, pp.763-770.
H. Aymeric, and C.M. Christine, Detection robuste et automatique des con- T.F. Cootes and C.J. Taylor, Combining point distribution models with shape
tours myocardiques sur des sequences IRM cardiaques marquees, GRETSI models based on finite element analysis, BMVC94 2 (1994), 419428.
(2005).
H. Delingette and J. Montagnat, Shape and topology constraints on paramet-
K.O. Babalola, B. Patenaude, P. Aljabar, J. Schnabel, D. Kennedy, W. ric active contours, Comput Vis Image Understanding 83 (2001), 140171.
Crum, S. Smith, T. Cootes, M. Jenkinson, and D. Rueckert, An evaluation
B.C. Dickerson, E. Feczko, J.C. Augustinack, J. Pacheco, J.C. Morris, B.
of structures in the brain, Neuroimage 47 (2009), 14351414 47.
Fischl, and R.L. Buckner, Differential effects of aging and Alzheimers dis-
M. Bar and E. Aminoff, Cortical analysis of visual context, Neuron 38 ease on medial temporal lobe cortical thickness and surface area, Neurobiol
(2003), 347358. Aging 30 (2009), 432440.
B. Bascle. Contributions et applications des modeles deformables en vision H. M. Duvernoy. The Human Hippocampus: Functional Anatomy, Vascu-
par ordinateur. These de doctorat, universite de Nice-sophia antipolis, 1994. larization and Serial Sections with MRI. The Human Hippocampus, Spring-
M.F. Bear, F.W. Connors, and M.A. Paradiso. Chapitre 1 : Neurosciences - a er, 2005.
la decouvertes du cerveau. Edditions Pradel. 2007. S. Ettaieb, K. Hamrouni, and S. Ruan. Statistical models of shape and spa-
J. Bollinger, M.T. Rubens, E. Masangkay, J. Kalkstein, and A. Gazzaley, tial relation-application to hippocampus segmentation. 9th International
An expectation-based memory deficit in aging, Neuropsychologia 49 Conference on Computer Vision Theory and Applications (Visapp), Lisbon-
(2011), 14661475. Portugal, Janvier, 2014.
R.L. Buckner, J.R. Andrews-Hanna, and D.L. Schacter, The brains default R. Epstein, A. Harris, D. Stanley, and N. Kanwisher, The parahippocampal
network: anatomy, function, and relevance to disease, Ann N Y Acad Sci 11 place area: recognition, navigation, or encoding, Neuron 23 (1999), 115125.
(2008), 138. M Kayalvizhi, G Kavitha, and CM Sujatha, Analysis of ventricle regions in
L. Caselles, F. Catt, C. Coll et, and F. Dibos, A geometric model for active Alzheimers brain MR images using level set based methods, International
contours in image processing, Numer Math (1993), 131. Journal of Biomedical Engineering and Technology 12, 300319.

Vol. 27, 311 (2017) 11