Unsupervised Method based on Probabilistic Neural Network for the
Segmentation of Corpus Callosum in MRI Scans
Amal Jlassi
1
, Khaoula ElBedoui
1,2
, Walid Barhoumi
1,2
and Chokri Maktouf
3
1
Universit
´
e de Tunis El Manar, Institut Sup
´
erieur d’Informatique, Research Team on Intelligent Systems in Imaging and
Artificial Vision (SIIVA), LR16ES06 Laboratoire de recherche en Informatique, Mod
´
elisation et Traitement de l’Information
et de la Connaissance (LIMTIC), 2 Rue Bayrouni, 2080 Ariana, Tunisia
2
Universit
´
e de Carthage, Ecole Nationale d’Ing
´
enieurs de Carthage, 45 Rue des Entrepreneurs, 2035 Carthage, Tunisia
3
Biophysics and Nuclear Medicine Department, Pasteur Institute of Tunis, 13 Place Pasteur, 1002 Tunis, Tunisia
chokri.maktouf@rns.tn
Keywords:
Corpus Callosum, MRI, Unsupervised Classification, Probabilistic Neural Network, Cluster Validity Index.
Abstract:
In this paper, we introduce an unsupervised method for the segmentation of the Corpus Callosum (CC) from
Magnetic Resonance Imaging (MRI) scans. In fact, in order to extract the CC from sagittal scans in brain
MRI, we adopted the Probabilistic Neural Network (PNN) as a clustering technique. Then, we used k-means
to obtain the target classes. After that, we introduced a cluster validity measure based on the maximum entropy
principle (Vmep), which aims to define dynamically the optimal number of classes. The later criterion was
applied in the hidden layer output of the PNN, while varying the number of classes. Finally, we isolated the
CC using a spatial-based process. We validated the performance of the proposed method on two challenging
datasets using objective metrics (accuracy, sensitivity, Dice coefficient, specificity and Jaccard similarity), and
the obtained results proved the superiority of this method against relevant methods from the state of the art.
1 INTRODUCTION
Advances in imaging techniques during the past de-
cade, especially in modalities of magnetic resonance,
made it easy for neuroscientists and clinicians to
study the Corpus Callosum (CC) in depth. Many of
these studies focused on analyzing the correlation be-
tween the CC’s dimensions and some neurological
diseases. In fact, the CC is the largest white mat-
ter structure and the biggest fiber tract within the hu-
man brain that is responsible for the communication
between the two cerebral hemispheres (Wong et al.,
2006) (Waxman, 2003). It transmits visual, moto-
ric, somatosensory, and auditory information from
one hemisphere to another (Waxman, 2003) (Ganjavi
et al., 2011). However, the CC shape might be the
cause of some diseases such as epilepsy, Alzheimer
and other types of psychosis. In a recent study, 42 pa-
tients with temporal lobe epilepsy and Hippocampal
Sclerosis (HS) along with 30 subjects were studied
with Diffusion Tensor Imaging (DTI) to evaluate the
integrity of the CC (Lyra et al., 2017). Results sho-
wed that some clinical characteristics; like seizure fre-
quency, duration and lesions; are associated with ab-
normalities in the CC. Furthermore, deformities in the
CC shape had been observed in several neurodegene-
rative diseases such as childhood stuttering and smo-
king, which seemed to influence the CC shape (Choo
et al., 2012) (Choi et al., 2010). Another study (Prigge
et al., 2013) conducted on 917 individuals confirms
that CC abnormalities might be the cause of autism.
Indeed, some hypothesis implies that autism is caused
by a defect in the CC which is probably related to the
number of white fibers responsible for the communi-
cation between the two cerebral hemispheres. In any
case, studies results can be used to predict future cases
of these diseases (Lyra et al., 2017). Thus, the goal of
magnetic resonance brain imaging in the framework
of CC diagnosis is to unfold neurological patterns in
the development of different diseases in order to en-
hance the related treatment (Lainhart et al., 2002).
Nevertheless, the visual inspection of CC structures
in MRI scans suffers from intra-variability and inter-
variability between clinicians, even for experienced
ones. In fact, the CC area in sagittal brain MRI sli-
ces is generally composed of different small regions
forming a narrow, horizontally oriented shape which
is generally located near the center of the brain (Fi-
Jlassi, A., ElBedoui, K., Barhoumi, W. and Maktouf, C.
Unsupervised Method based on Probabilistic Neural Network for the Segmentation of Corpus Callosum in MRI Scans.
DOI: 10.5220/0007400205450552
In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2019), pages 545-552
ISBN: 978-989-758-354-4
Copyright
c
2019 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
545
gure 1.a). However, in many sagittal brain MRI sli-
ces, the fornix appears in the neighborhood of the CC
with a similar intensity (Figure 1.b). Hence, there has
been a growing need for Computer-Aided Diagnosis
(CAD) systems for automated analysis of the corpus
callosum in MRI scans. Indeed, the subdivision of
the CC is highly recommended for the parcellation
task. This fact makes it possible to more effectively
study brain shape and connectivity and thus to ana-
lyze the properties inside the structure (Park et al.,
2008). Besides, the shape and location of CC play
a very important role in identifying the brain from ot-
her tissues. In this context, various works have been
focused on segmenting automatically the CC, given
the strong impact of the segmentation quality on the
overall precision of the CAD of the CC. Therefore,
since the area of the CC is characterized by a high
intensity, a precise segmentation of CC within brain
MRI scans, without penetrating the irrelevant neig-
hboring structures, is a challenging task in the diag-
nosis and treatment processes. To deal with this issue,
we propose to automatically delimit the CC within
MRI images. The contribution of this work is two-
fold. (1) As best as we know, we adopted for the first
time the Probabilistic Neural Network (PNN) for the
segmentation of CC (Cover et al., 2018). This co-
mes from the promising results recorded by the Neu-
ral Networks (NN) within classification tasks. In fact,
the most commonly used NN are radial function net-
works, and notably the PNN which proved its effi-
ciency in many applications (Zhang, 2000). (2) Besi-
des, the learning process of the proposed unsupervi-
sed PNN-based method is fully unsupervised, with no
parameter adjusting and instantaneous training.
(a) (b)
Figure 1: Examples of sagittal brain MRI slices: (a) CC area
is composed of rostrum (in red), genu (in orange), anterior
body (in yellow), mid-body (in green), posterior body (in
blue), isthmus (in purple) and splenium (in light purple). (b)
An example where the fornix (framed in red) appears in the
neighborhood of the CC while being of similar appearance.
The rest of this paper is organized as follows. In
Section 2, we discuss the related work. Then, we des-
cribe the proposed method in Section 3. We show
results in Section 4. Finally, a conclusion with some
directions for future work is presented in Section 5.
2 RELATED WORK
Various 2D segmentation methods were proposed to
improve the precision of the CC detection. Howe-
ver, the most of these methods have not overcome all
challenges encountered. In fact, the CC segmenta-
tion is a challenging task given that a normal shape
of the CC might not clearly highlight internal de-
formities, what can add complexity to the diagnosis
process. Besides, many internal abnormalities might
include bumps, which are hard to detect when per-
forming CC segmentation. Existing CC segmenta-
tion methods can be regrouped into two main clas-
ses: supervised methods and unsupervised ones. On
the one hand, within the class of supervised methods,
a CC delineation method based on the atlas appro-
ach was developed in (Ardekani et al., 2012), where
authors applied a rectangular CC search area based
on a priori database information. This method con-
sists to compute the local cross-correlation map where
pixel values are represented by the cross-correlation
between the warped atlas and the test image. After
that, the association of a pixel to the CC region is
performed by selecting the pixel having the highest
cross-correlation, and a specified vote rule is applied
to merge the corresponding classifications. However,
this method depends on a priori information from the
atlas dataset. Differently, in order to segment the CC
structure, (Divya and Vishnu, 2014) proposed to start
by locating coarsely the CC area. This is can be per-
formed by the adaptive mean shift algorithm or the
k-means clustering. Once the CC area is located, it
is used as the initial contour for the geometric active
contour. This method accuracy depends strongly on
the initialization precision of the CC region, so an er-
ror initialization of the CC region can lead to an incor-
rect segmentation. Moreover, (Farhangi et al., 2016)
proposed to embed shape information into level set
image segmentation. Indeed, the CC segmentation
was based on inferring shape variations by a linear
combination of instances in the shape repository. This
allows the guidance of segmenting curve towards the
boundary of an object as well as the conservation con-
sistent with the shapes provided in the training set.
Indeed, a shift of the evolution curve at each step is
done in order to minimize the Chan-Vese energy as
well as towards the best approximation based on a li-
near combination of learning samples. This supervi-
sed method is effective only with a sufficient number
of training shapes and a linear combination of lear-
ning sets. Generally, it remains true that supervised
methods are conceptually simple but their robustness
is strongly dependent on the accuracy of their train-
ing process. On the other hand, within the class of
VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications
546
unsupervised methods, a fully automated hybrid met-
hod was proposed in (Li et al., 2017). This method is
based on an Adaptive Mean Shift (AMS) clustering.
Indeed, to identify the CC region, an automatic CC
contour initialization is performed while applying an
area analysis, followed by matching template based
on the shape, and locating analysis. Then, the Geo-
metric Active Contour (GAC) model was used for the
delineation of the boundaries of detected areas. Ne-
vertheless, the application of the AMS may affect sur-
rounding tissues to the CC cluster what causes unsa-
tisfactory segmentation results. Thus, this method re-
mains valid just for some specific images. In (
˙
Ic¸er,
2013), the author presented a comparative study of
Gaussian Mixture Modeling (GMM) and Fuzzy C-
Means (FCM) methods for the segmentation of CC.
The GMM is used to define image classes while using
the weighted sum of Gaussian distributions and ap-
plying statistical decisions. By using FCM, image
classes are represented by membership functions ac-
cording to fuzziness information that expresses the
distance from the cluster centers. Then, a maximum
clustering is used to achieve the final segmentation. A
fully automated method for the segmentation of CC
in brain MRIs was proposed in (Tang et al., 2016). In
this method, it is supported that a supervised approach
can lead to a breakthrough in CC segmentation perfor-
mance compared to many unsupervised segmentation
methods. In order to automatically learn a set of la-
tent features that are useful for identifying the target
structures, a discriminatory learning framework was
used. In the first step, a multi-atlas-based segmenta-
tion approach is used for localizing the CC structure
of each image. The second step consists in the joint
feature-learning and model training using a Convo-
lutional Encoder Network (CEN). Nevertheless, this
method requires a post-treatment step to reduce the
amount of false positive predictions. In summary,
unsupervised methods are more complex than super-
vised methods but they have an undeniable advantage
in that they do not require prior knowledge.
3 PROPOSED METHOD
The proposed method is composed of three steps:
image preprocessing using Anisotropic Diffusion Fil-
tering (ADF), classification based on unsupervised
PNN, and CC isolation using a spatial filtering.
3.1 Anisotropic Diffusion Filter
To have a good standard of the MR brain images,
a preprocessing step, which allows experts or ima-
ging modalities an efficiency of further processing,
is generally required. This step aims to improve the
signal-to-noise ratio, to remove undesired parts in the
background, and to smooth the inner part of the re-
gion while preserving its edges. These parameters
are usually influenced by other parameters such as
patient comfort and physiological constraints (Demi-
rhan et al., 2015). Therefore, we applied the ADF in
order to improve the clarity of the CC within the input
raw MRI scan. In fact, to deblur high-frequency noise
while preserving the main edges of existing objects,
the use of ADF (Perona and Malik, 1990) has proven
effective. Indeed, based on its confirmed advantages,
the ADF preprocessing method is frequently used in
the context of MR imaging (Palma et al., 2014). The
ADF uses a diffusion coefficient as an edge detector in
order to obtain a smoothed image with preserved ed-
ges. In fact, to obtain the output image, we followed
an iterative process based on the following equation
(Kesareva, 2017):
I
0(k+1)
(i, j) =
I(i, j) + λg(i, j)
wε(i, j)
I
0(k)
(w)
1 + λg(i, j)
|
ε
|
, (1)
where, I is the input image, I
0
is the resultant
image after the k
th
iteration, g (=
1
1+
|
I
|
2
) is an edge
indicator function, λ is the regularization parameter
that defines the trade-off between removing noise and
preserving sharp boundaries, ε(i, j) denotes the neig-
hborhood of the pixel (i,j), and
|
.
|
denotes the set car-
dinality operator. Hence, our ADF-based preproces-
sing, using three iterations, allows to minimize the
presence of undesired contours (Figure 2) (e.g. for-
nix contour) what optimize the overall performance
of the proposed CC segmentation method.
(a) (b)
Figure 2: Preprocessing using ADF: (a) Original image. (b)
Filtered image.
3.2 Classification
The classification step is based on the Probabilis-
tic Neural Network (PNN). In fact, this feed-forward
neural network is used almost for classification, seen
its advantage of getting better results for the Bayes-
based Decision under certain conditions and its ro-
Unsupervised Method based on Probabilistic Neural Network for the Segmentation of Corpus Callosum in MRI Scans
547
bustness against noise (Georgiadis et al., 2008). The
PNN is made up of three main layers: an input layer,
a hidden layer and an output layer (Shree and Kumar,
2018). In the first step, the PNN receives iteratively
a D-dimensional feature vector x = (x
1
,...,x
D
) for the
input neurons x
i
(1 i D), where D denotes the
size of the input image. This vector is transmitted to
the neurons in the hidden layer which is composed of
nodes that are gathered in groups (each of the C clas-
ses belongs to a group). In fact, a centred Gaussian
function f (x) is associated with each hidden node in
the class k (1 k C), what defines the Probability
Density Function (PDF) (2).
f
k
(x) =
1
(2π)
D/2
σ
D
e
((
k
xx
k
k
2
/
(
2σ
2
)))
, (2)
where, σ is the smoothing parameter for the Gaussian,
D is the dimension of the input vector x and
k
x x
k
k
is the Euclidean distance between the vector x and the
centre x
k
of the k
th
cluster. At the level of the second
layer, a summation of the contribution for each class
of inputs is made and its total output, as a vector of
probabilities, is estilatmed as follows:
P
k
(x) =
1
(2π)
D/2
σ
D
C
C
k=1
e
((
k
xx
k
k
2
/
(
2σ
2
)))
, (3)
where C is the number of output nodes. Finally, a
competitive transfer function attributes 1 for the input
class that has the maximum PDF join and 0 for ot-
her classes. Therefore, Bayes decision rule under the
following assumption is used to determine the class
belongingness of the output x in the decision layer:
c(x) = argmax
{
P
k
(x)
}
,k = 1,2,...,C, (4)
where c(x) is the estimated class of the output
x.Accordingly, given the input image I and the in-
terval [C
min
: C
max
] of possible values of the number
of classes in I, the unfolding stages of the proposed
method, which outputs a mask M
CC
that isolates the
CC in the input image I, can be summarized as in
Algorithm 1 (such that N = Cmax + 1 Cmin). In
other words, the proposed classification step consists
in defining the target classes (Cl f
N
) using k-means,
then applying the PNN for the classification accor-
ding to these classes. In addition, we used a cluster
validity function that dynamically allows optimizing
the choice of the number of classes in a given inter-
val. In fact, for the PNN, defining the modes (centers
of the Gaussian functions) is a crucial task. Then, an
evaluation method, called the cluster validity, is even
necessary to determine the optimal number of clusters
C. In order to automate the PNN, a summation of the
probability density functions in the output of its hid-
den layer was made. It takes the form of a matrix of
probabilities (PM) representing the memberships of
pixels x
k
to the classes C
i
. This matrix will be used to
calculate the Validity Index V ld by varying the class
number C in a [C
min
: C
max
] (C
min
and C
max
denote the
minimum and the maximum number of possible clas-
ses, respectively) defined by the user. When the va-
lidity index V ld reaches its maximum value, we get
the optimal number of classes that will be adopted for
the classification process based also on PNN. To deal
with this challenging task, the output of this PNN-
based classification is a cluster map (Figure 3).
Algorithm 1: Proposed CC segmentation method.
Input: I, C
min
, C
max
Output: M
CC
I
0
ADF(I)
for C = C
min
: C
max
do
Cl f
1
...Cl f
N
kmeans(I
0
,C)
end for
for H = 1 : N do
PM
H
PNN(I
0
,Cl
H
)
end for
Max 0
for k = 1 : N do
if V
mep
(PM
k
) > Max then
Max V
mep
(PM
k
)
C
k
end if
end for
ClusterMap PNN(I
0
,C
)
M
CC
Isolation(ClusterMap)
(a) (b)
Figure 3: Classification using PNN: (a) An input sagittal
brain MRI. (b) Resulting cluster map generated by PNN.
In particular, we adopted the k-means method to
obtain Gaussian functions centers in the hidden layer.
In fact, in order to reduce the overlap of the centers,
we used a spread that is equal to half of the minimum
distance between the neighboring centers, in order to
locally determine the widths of the radial basis functi-
ons. Indeed, the aim of clustering is to identify groups
of similar pixels and thus discover an interesting dis-
tribution of model. The majority of clustering algo-
rithms require knowledge of the right number of C
classes to have an efficient classification. We are per-
forming obliged to measure the performance of the
VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications
548
classification by using a validity criterion. This makes
it possible to choose the optimal partition among all
those obtained with the various plausible number of
clusters. After a comparative study concerning clus-
ter validity functions, we opted to use V mep index
(5) which is based on the maximum entropy principle
(Ammor et al., 2008) which is an automatic approach
that does not require any parameter settings.
V ld = V mep(PM) =
1
k
k
j=1
S
j
+ ln(k), (5)
where, 1 k Cl
H
and S
j
=
iCl
H
PM
i j
ln(PM
i j
).
The details of the proposed method for the automated
PNN-based classification are summarized in Figure 4.
Once the CC class is identified, a spatial-based filte-
ring is applied in order to isolate the CC region. Thus,
the output of the proposed method is a binary mask of
the M
CC
(0 for the background and 1 for the M
CC
).
Then, in order to define the CC contour, a follow-up
algorithm applied on the border pixels of the CC re-
gion that are characterized by a maximum of the spa-
tial gradient (Barhoumi et al., 2002).
Figure 4: Outline of the proposed PNN-based classification.
4 EXPERIMENTAL RESULTS
For the evaluation of the proposed method, we used
brain MRI scans from two datasets. On the one hand,
we used the Open Access Series of Imaging Stu-
dies (OASIS) dataset, which is publicly available on
www.oasis-brains.org. Each MR image within this
dataset is composed of 128 slices with a resolution of
256 × 256 (1 × 1 mm). This dataset includes a cross-
sectional collection of 416 subjects with 420 MR ima-
ges, such that the majority of images are sagittal secti-
ons, for men and women aged from 18 to 96 years.
The subjects are all right-handed and include indivi-
duals with early-stage Alzheimers Disease (AD). All
images were acquired on a 1.5-T Vision scanner (Sie-
mens) and using identical sequences in a single ima-
ging session. It contains images qualified by a qua-
lity control since it excludes from the distribution of
any image with severe artifacts. On the other hand,
we used the MRI DICOM dataset of the head of a
male professor, from the Radiology Department at the
Macclesfield General Hospital, who is aged 52 years.
The subject suffers from a small vertical strabismus
(hypertropia), a misalignment of the eyes, which is
visible in the dataset. The MRI scans within this data-
set are T2 weighted turbo-spin-echo (T2W TSE) and
T1 weighted Fast Field Echo (T1W FFE) with a re-
solution of 512 × 512. It is worthy noting that these
datasets offer different sections (Axial section, Coro-
nal section, and Sagittal section) what allows to study
several pathologies effectively. However, clinicians
usually examine the form and/or the size of the CC by
visually interpreting just the sagittal sections. Thus,
we used these sections to obtain accurate results on
the form of CC since they allow a better analysis and
diagnosis of CC-related diseases.
4.1 Qualitative Evaluation
Figure 5 shows the visual assessment of the recorded
results for a sample of challenging brain MRI scans,
what confirms the accuracy of the proposed method
for the CC segmentation. In fact, according to our col-
laborator clinician expert, the CC shape and thickness
are well defined and the delineated CC area shows
closely the four anatomical divisions of the CC, es-
pecially the critical ones, notably the rostrum and the
splenium. Furthermore, the fornix is correctly remo-
ved from the CC area. Indeed, the obtained CC masks
show a precise segmentation of CC within brain MRI
scans, without penetrating the irrelevant neighboring
structures. Note that, within the selected sample of
MRI brain scans, the CC is extracted both on female
(lines 1 and 3) and male (lines 2 and 4) subjects. It is
also delineated on demented (line 2) as well as non-
demented subjects. Moreover, the CC extraction was
performed on important (line 3) and normal intracra-
nial volume. Moreover, Figure 6 shows an example
of an incorrect segmentation, while using a relevant
method for the state of the art (Li et al., 2017), of
the CC. In fact, while using this method, the neighbo-
ring tissues, whose have a similar intensity to the CC,
were mis-segmented as a part of the CC. However,
the proposed method has successfully solved this pro-
blem within this MRI example, by recording accurate
results that separate the connected regions from CC.
Unsupervised Method based on Probabilistic Neural Network for the Segmentation of Corpus Callosum in MRI Scans
549
Row 1Row 1Row 1Row 1
Figure 5: Qualitative evaluation: (a) Input image. (b) Cluster Map. (c) Isolated CC. (d) Ground-truth (e) Delineation of the
CC bythe proposed method (except the two last MRI scans, where C is equal to 4, C is equal to 5 for the remaining scans).
Concerning the parameter tuning, it is important to
notice that we set C
min
and C
max
to 2 and 8, respecti-
vely. This choice was motivated by the fact that this
range is essential for good clustering performance as
well as for the minimization of the execution time,
since the MRI brain scans generally contains between
4 and 6 anatomical classes.
4.2 Quantitative Evaluation
For the quantitative evaluation of the proposed met-
hod comparatively to other relevant methods from
the state of the art, we used, as evaluation metrics,
the accuracy, the sensitivity, the Dice coefficient, the
specificity and the Jaccard similarity. The expressi-
ons of these standard metrics are defined in Table 1,
where TP refers to the True Positive (image region
correctly classified as CC), TN refers to the True Ne-
gative (image region which is correctly classified as
background), FP refers to the False Positive (image
region which is incorrectly classified as CC) and fi-
nally FN refers to the False Negative (image region
(a)
(b)
(c)
Figure 6: Comparison of the proposed method against a re-
levant AMS-based method (Li et al., 2017): (a) Input scan
from the OASIS dataset. (b) Incorrect classification (for-
nix is classified as a part of the CC) using the AMS-based
method. (c) Accurate classification using the proposed met-
hod (the maximum validity index for this scan was equal to
0.951, which corresponds to C equals to 5).
which is incorrectly classified as background). We
notice that we produced, for the first time, a very use-
ful ground-truth for CC segmentation within the chal-
lenging widely used OASIS dataset. In fact, a profes-
VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications
550
Table 1: Evaluation metrics.
Metrics Expression Description
Accuracy
(T P+T N)
(T P+FN+T N+FP)
It is presented as the rate of correctly classified items.
Sensitivity
T P
(T P+FN)
It refers to the proportion of positive items correctly classified.
Dice coefficient
2×T P
2×T P+FN+FP
It is defined as a statistical measure that is used for comparing
the similarity of two sample sets.
Specificity
T N
(T N+F P)
It is the rate of negative items rightly identified.
Jaccard Similarity
T P
T P+FN+FP
It measures similarity between two sample sets.
Table 2: Evaluation of the proposed method comparatively to the other segmentation methods (Best value are on bold).
Accuracy Jaccard Simila-
rity
Sensitivity Specificity Dice coeffi-
cient
(Li et al., 2017) 0.95 ± 0.09 0.84 ± 0.08
(
˙
Ic¸er, 2013) 0.983 ± 0.0072 0.966 ± 0.0083 0.966 0.97
(
˙
Ic¸er, 2013) 0.97 ± 0.008 0.942 ± 0.0092 0.986 0.934
(Tang et al., 2016) 0.91 ± 0.02
(Farhangi et al.,
2016)
0.93
Proposed method 0.99 ± 0.005 0.99 ± 0.004 0.94 ± 0.149 0.98 ± 0.13 0.99 ± 0.008
sional neurologist from Pasteur Institute of Tunis has
been charged with manually drawing the CC regions
from all images belonging to the OASIS dataset. Be-
sides, we applied a post-processing in order to exclu-
sively extract the CC area. Table 2 shows the recorded
results, where we produce the obtained performances
by presenting the mean value and the standard devi-
ation of the used objective metric. In fact, we pro-
duced the recorded metric for the proposed method
as well as for five relevant CC segmentation methods
from the state of the art. The compared methods are:
AMS ACI GAC (Li et al., 2017); which combi-
nes Adaptive Mean Shift technique (AMS), Automa-
ted Initialization of CC Contour (ACI) and Geome-
tric Active Contour model (GAC); Gaussian Mixture
Model (GMM) (
˙
Ic¸er, 2013), Fuzzy C-Means (FCM)
(
˙
Ic¸er, 2013), ’ISP into AC’ that is based on Incorpo-
rating Shape Prior (ISP) into Active Contours (AC)
with a Sparse Linear Combination of Training Shapes
(Tang et al., 2016), RT JSFE that uses Robust
Target-Localization (RTL) and Joint Supervised Fea-
ture Extraction and Prediction (JSFE) (Farhangi et al.,
2016). It is clear that the proposed method records the
best Jaccard similarity score (= 0.99 ± 0.004) compa-
ratively to the compared methods. We can conclude
that the proposed method localizes precisely and deli-
mits robustly the CC (average of approximately 0.99
with 0.004 as standard deviation). This low standard
deviation reflects a low dispersion of the analyzed va-
lues what it increases considerably the accuracy rate
of the proposed method. Evenly, it reaches the best
Dice coefficient, accuracy and specificity scores with
0.99 ± 0.008, 0.99 ± 0.005 and 0.98 ± 0.13, respecti-
vely. On the other hand, for sensitivity metric, the
proposed method reaches good score and still better
than AMS ACI GAC method. The decline of the
proposed method performance according to this me-
tric can be explained by the cause of the ground-truth
which is manually drawing.
5 CONCLUSION
CC is the largest white matter structure and the big-
gest fiber tract within the human mind that it is re-
sponsible for the communication between the two ce-
rebral hemispheres. The CC shape might be the cause
of some diseases such as autism, epilepsy and Alz-
heimer. Thus, the segmentation of the CC from MRI
images can be used to predict future cases of diseases
or to unfold neurological patterns in the development
of different diseases. To deal with these challenges,
we introduced an unsupervised segmentation method,
which is based on the automation of the PNN. This
method has been extensively validated on two chal-
lenging standard datasets. Indeed, the proposed met-
hod achieved higher performance values than relevant
methods from the state of the art. Indeed, with the
suggested method, the Dice coecient reached 99%,
whereas specicity reached 98% and sensitivity rea-
ched 94%. Furthermore, the proposed method can be
used to study the CC morphometry which is important
in the diagnosis of many neurological diseases.
Unsupervised Method based on Probabilistic Neural Network for the Segmentation of Corpus Callosum in MRI Scans
551
REFERENCES
Ammor, O., Lachkar, A., Slaoui, K., and Rais, N. (2008).
Optimal fuzzy clustering in overlapping clusters. In-
ternational Arab Journal of Information Technology
(IAJIT), 5(4).
Ardekani, B. A., Toshikazu, I., Bachman, A., and Szeszko,
P. R. (2012). Multi-atlas corpus callosum segmenta-
tion with adaptive atlas selection.
Barhoumi, W., Zagrouba, E., Damiani, E., Howlett, R., and
Jain, L. (2002). Boundaries detection based on poly-
gonal approximation by genetic algorithms. Frontiers
in Artificial Intelligence and Applications. New York:
Springer, pages 1529–33.
Choi, M.-H., Lee, S.-J., Yang, J.-W., Kim, J.-H., Choi, J.-S.,
Park, J.-Y., Jun, J.-H., Tack, G.-R., Lee, B.-Y., Kim,
H.-J., et al. (2010). Difference between smokers and
non-smokers in the corpus callosum volume. Neuros-
cience letters, 485(1):71–73.
Choo, A. L., Chang, S.-E., Zengin-Bolatkale, H., Ambrose,
N. G., and Loucks, T. M. (2012). Corpus callosum
morphology in children who stutter. Journal of com-
munication disorders, 45(4):279–289.
Cover, G., Herrera, W., Bento, M. P., Appenzeller, S., and
Rittner, L. (2018). Computational methods for corpus
callosum segmentation on mri: A systematic literature
review. Computer methods and programs in biomedi-
cine, 154:25–35.
Demirhan, A., T
¨
or
¨
u, M., and G
¨
uler, I. (2015). Segmentation
of tumor and edema along with healthy tissues of brain
using wavelets and neural networks. IEEE journal of
biomedical and health informatics, 19(4):1451–1458.
Divya, Manasa, M. and Vishnu, Priya, T. (2014). A hybrid
technique for the automated segmentation of corpus
callosum in midsagittal brain mri. Journal of Engi-
neering Research and Applications, 4(8):1–4.
Farhangi, M. M., Frigui, H., Bert, R., and Amini, A. A.
(2016). Incorporating shape prior into active contours
with a sparse linear combination of training shapes:
Application to corpus callosum segmentation. pages
6449–6452.
Ganjavi, H., Lewis, J. D., Bellec, P., MacDonald, P. A., Wa-
ber, D. P., Evans, A. C., Karama, S., Group, B. D. C.,
et al. (2011). Negative associations between corpus
callosum midsagittal area and iq in a representative
sample of healthy children and adolescents. PLoS
One, 6(5):e19698.
Georgiadis, P., Cavouras, D., Kalatzis, I., Daskalakis, A.,
Kagadis, G. C., Sifaki, K., Malamas, M., Nikifori-
dis, G., and Solomou, E. (2008). Improving brain
tumor characterization on mri by probabilistic neural
networks and non-linear transformation of textural fe-
atures. Computer methods and programs in biomedi-
cine, 89(1):24–32.
˙
Ic¸er, S. (2013). Automatic segmentation of corpus collasum
using gaussian mixture modeling and fuzzy c means
methods. Computer methods and programs in biome-
dicine, 112(1):38–46.
Kesareva, E. (2017). An algorithm for estimation an aniso-
tropic diffusion filter parameter. pages 682–685.
Lainhart, J. E., Ozonoff, S., Coon, H., Krasny, L., Dinh, E.,
Nice, J., and McMahon, W. (2002). Autism, regres-
sion, and the broader autism phenotype. American
Journal of Medical Genetics, 113(3):231–237.
Li, Y., Wang, H., Ahmed, N., and Mandal, M. (2017). Auto-
mated segmentation of corpus callosum in midsagittal
brain mris. ICTACT JOURNAL ON IMAGE AND VI-
DEO PROCESSING, 8(1):1554–1565.
Lyra, K. P., Chaim, K. T., Leite, C. C., Park, E. J., Andrade,
C. S., Passarelli, V., Val
´
erio, R. M., Jorge, C. L., Cas-
tro, L. H., and Otaduy, M. C. (2017). Corpus callosum
diffusion abnormalities in refractory epilepsy associ-
ated with hippocampal sclerosis. Epilepsy research,
137:112–118.
Palma, C. A., Cappabianco, F. A., Ide, J. S., and Miranda,
P. A. (2014). Anisotropic diffusion filtering operation
and limitations-magnetic resonance imaging evalua-
tion. IFAC Proceedings Volumes, 47(3):3887–3892.
Park, H.-J., Kim, J. J., Lee, S.-K., Seok, J. H., Chun, J.,
Kim, D. I., and Lee, J. D. (2008). Corpus callosal
connection mapping using cortical gray matter parcel-
lation and dt-mri. Human brain mapping, 29(5):503–
516.
Perona, P. and Malik, J. (1990). Scale-space and edge
detection using anisotropic diffusion. IEEE Tran-
sactions on pattern analysis and machine intelligence,
12(7):629–639.
Prigge, M. B., Lange, N., Bigler, E. D., Merkley, T. L.,
Neeley, E. S., Abildskov, T. J., Froehlich, A. L., Niel-
sen, J. A., Cooperrider, J. R., Cariello, A. N., et al.
(2013). Corpus callosum area in children and adults
with autism. Research in autism spectrum disorders,
7(2):221–234.
Shree, N. V. and Kumar, T. (2018). Identification and clas-
sification of brain tumor mri images with feature ex-
traction using dwt and probabilistic neural network.
Brain informatics, 5(1):23–30.
Tang, L. Y., Brosch, T., Liu, X., Yoo, Y., Traboulsee, A.,
Li, D., and Tam, R. (2016). Corpus callosum segmen-
tation in brain mris via robust target-localization and
joint supervised feature extraction and prediction. pa-
ges 406–414.
Waxman, S. (2003). Lange clinical neuroanatomy. New
York: McGraw-Hill, 27nd edition.
Wong, T.-T., Kwan, S.-Y., Chang, K.-P., Hsiu-Mei, W.,
Yang, T.-F., Chen, Y.-S., and Yi-Yen, L. (2006). Cor-
pus callosotomy in children. Child’s Nervous System,
22(8):999–1011.
Zhang, G. P. (2000). Neural networks for classifica-
tion: a survey. IEEE Transactions on Systems, Man,
and Cybernetics, Part C (Applications and Reviews),
30(4):451–462.
VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications
552