Biomedical Signal Processing and Control 39 (2018) 64–73

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

journal homepage: www.elsevier.com/locate/bspc

Dental diagnosis from X-Ray images: An expert system based on

fuzzy computing
Le Hoang Son a,∗ , Tran Manh Tuan b , Hamido Fujita c , Nilanjan Dey d , Amira S. Ashour e ,
Vo Truong Nhu Ngoc f , Le Quynh Anh f , Dinh-Toi Chu g,h
a
VNU University of Science, Vietnam National University, Vietnam
b
Thuy Loi University, Vietnam
c
Iwate Prefectural University, Iwate, Japan
d
Techno India College of Technology, Kolkata, India
e
Faculty of Engineering, Tanta University, Egypt
f
School of Odonto - Stomatology, Hanoi Medical University, Hanoi, Vietnam
g
Institute for Research and Development, Duy Tan University, 03 Quang Trung, Danang, Vietnam
h
Faculty of Biology, Hanoi National University of Education, Hanoi, Vietnam

a r t i c l e i n f o a b s t r a c t

Article history: Background: Computerized medical diagnosis systems from X-Ray images are of great interest to physi-
Received 2 April 2017 cians for accurate decision making of possible diseases and treatments. Subclinical disease has no
Received in revised form 12 June 2017 recognizable clinical findings, thus it is desirable to segment the dental X-Ray image into groups and
Accepted 11 July 2017
then use soft computing methods to check the possibility of whether or not any disease occurs therein.
Methods: The current work proposed a novel framework called Dental Diagnosis System (DDS) for dental
Keywords:
diagnosis based on the hybrid approach of segmentation, classification and decision making. It utilized the
Dental X-Ray image
best dental image segmentation method based on semi-supervised fuzzy clustering for the segmentation
Decision making
Graph-based clustering
task. A new graph-based clustering algorithm called APC+ for the classification task was proposed. A new
Medical diagnosis decision making procedure was designed to determine the final disease from a group of diseases found
Semi-Supervised fuzzy clustering from the segments.
Results: The proposed DDS was modeled under the real dental case of Hanoi Medical University, Vietnam
including 87 dental images of five popular diseases, namely: root fracture, incluse teeth, decay, missing
teeth, and resorption of periodontal bone. The DDS accuracy is 92.74% which is superior to the other
methods namely fuzzy inference system (89.67%), fuzzy k-nearest neighbor (80.05%), prim spanning tree
(58.46%), kruskal spanning tree (58.46%), and affinity propagation clustering (90.01%).
Conclusion: Empirical results established that superior performance of the DDS to other related methods
the findings of the achieved results can assist dental clinicians in their professional work.
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction
Dental X-ray images analysis and processing play a significant
role for the dental diseases’/conditions’ diagnosis, treatment, and
studying the dental diseases nature as well as predicting dental
diseases at their early stage [28,29]. Typically, dental X-rays repre-
sent pictures of the teeth, soft tissues, and bones for determining
∗ Correspondence to: 334 Nguyen Trai, Thanh Xuan, Hanoi, Vietnam.
Tel.: +84 904 171 284.
E-mail addresses: sonlh@vnu.edu.vn (L.H. Son), tuan tm@tlu.edu.vn (T.M. Tuan),
HFujita-799@acm.org (H. Fujita), nilanjan.dey@tict.edu.in (N. Dey),
amirasashour@yahoo.com (A.S. Ashour), nhungoc@hmu.edu.vn (V.T.N. Ngoc),
lequynhanh@hmu.edu.vn (L.Q. Anh).
diseases/problems within the mouth, teeth, and jaw. It can dis-
play buried dental structures, cavities, and bone loss that visual
examination cannot detect [31].
In dental diagnosis, fuzzy inference system (FIS) is one of the
most common approaches for determining a projection from a
given input data set to an output data set using fuzzy logic [18,33].
Besides, Chattopadhyay et al. [7,8] presented an application of
Bayesian network to diagnose toothache. Kavitha et al. [23] used
Support Vector Machine (SVM) to predict the osteoporosis from
dental images. In 2014, Oad et al. [34] proposed a fuzzy-based
method to predict heart risk. Ramírez, Castillo and Soria [40,41]
used Fuzzy Neighbor K-Nearest Neighbor (FKNN) in different den-
tistry problems. Nevertheless, forming fuzzy rules and knowledge
in those works requires experts’ experience to avoid duplicate, con-

http://dx.doi.org/10.1016/j.bspc.2017.07.005
1746-8094/© 2017 Elsevier Ltd. All rights reserved.
 L.H. Son et al. / Biomedical Signal Processing and Control 39 (2018) 64–73 65

Fig. 1. Examples of dental X-ray images in each disease group: (a) Root fracture; (b) Incluse teeth; (c) Decay; (d) Missing teeth; (e) Resorption of periodontal bone.

flict and meaningless rules as well as guarantee accurate diagnosis Table 1

Statistical information of the groups of patients.
[16]. Such problems can be solved using a combination of seg-
mentation, classification and decision making algorithms, which Group Sex Age
can reduce the ambiguity and vagueness of determining the rules
Male Female 16–22 23–30 31–38
[11,37,38,42,45].
Root fracture 8 8 0 5 11
The first step of this process is to segment a dental X-Ray image
Incluse teeth 9 10 5 6 8
into groups for further examination of whether or not any disease Decay 9 8 8 3 6
exists. Dental X-Ray image segmentation is applied to create sev- Missing teeth 8 8 5 6 5
eral distinct groups in the image, whereas pixels in a group have Resorption of 10 9 4 7 8
more similarity than those in other groups [54]. This problem has periodontal bone
Sum 44 43 22 27 38
been studied extensively in [2–4,9,15,20,26,27,32,35,39,49–52,55].
The typical approach- semi-supervised algorithm uses additional
information, for example Bouchachia and Pedryzc [3] used the
task was proposed. A new decision making procedure was designed
membership matrix in the semi-supervised fuzzy c-mean algo-
to determine the final disease from a group of diseases found from
rithm (SSFCMBP). Yasunori et al. [49] carried out semi-supervised
the segments. In the dental professional view, the DDS was mod-
algorithm with standard fuzzy clustering (SSSFC) based on mix-
eled under the real dental cases including real dental images of
ing the membership function into the entire clustering process.
diseases’ classes [19].
Yin et al. [51] proposed semi-supervised entropy regularized fuzzy
The structure of the remaining sections is as follows. Section 2
clustering algorithm (eSFCM) that integrated the entropy factor
presents the materials and methods. Section 3 shows the experi-
into the semi-supervised clustering algorithm and used additional
ment results and discussions about performance of methods with
values to increase clustering performance. Portela et al. [36] pro-
further studies. Conclusions are highlighted in Section 4.
posed semi-supervised classifier based clustering algorithm for
magnetic resonance brain tissue segmentation.
Then in the next steps, once having the segments, a classifica- 2. Materials and methods
tion algorithm in [5,7,10,20,22,23,40,47] can be used to match the
segmented images with disease’ patterns in the database to con- 2.1. Dental images datasets
clude whether the image is for ‘disease’ or ‘non-disease’. Finally, a
decision making algorithm is applied to determine the final disease The proposed DDS was modeled under the real dental case
image from segments. of Hanoi Medical University, Vietnam including 87 dental images
In this paper, we propose a novel framework called Dental of five popular diseases in the period 2014–2015 [19]. The used
Diagnosis System (DDS) for dental diagnosis based on the hybrid dataset was collected from patients of age range from 16 to 38 and
approach of segmentation, classification and decision making. It divided into 5 groups, namely: i) 16 with root fracture, ii) 19 with
utilizes the best dental image segmentation method based on incluse teeth, iii) 17 with decay, iv) 16 with missing teeth, and v)
semi-supervised fuzzy clustering for the segmentation task. A new 19 with resorption of periodontal bone as shown in Table 1. Those
graph-based clustering algorithm called APC+ for the classification images taken by VATECH machine are both intraoral and panoramic
in each group (Fig. 1). Panoramic radiography captures the entire
66 L.H. Son et al. / Biomedical Signal Processing and Control 39 (2018) 64–73

mouth in a single image, including the teeth, upper and lower jaws, Edge-value and intensity in EEI measure the numbers of changes
surrounding structures and tissues [21]. of pixel values in a region:
In the current study, panoramic images were involved due to
their capability to provide coverage for periodontal bone defects, 
w/2

w/2

e (x, y) = b (x, y) (5)

periapical lesions, and pathological jaw lesions, and are better
p=−w/2q=−w/2
patient acceptance, less radiation exposure and infection control
[48]. Since the proposed DDS uses hybrid approach of segmen-

1, ∇ f (x, y) ≥ T1
tation, classification and decision making, thus it could deal with b (x, y) = (6)
other types of dental diseases such as malocclusion, pulpitis, peri- 0, ∇ f (x, y) < T1
odontitis, dental cyst, radicular cyst, and odontoma; which are  2  2
found to have clinical implications. Because the DDS is based on cal- ∂g (x, y) ∂g (x, y)
culating the similarity between a given image and existing disease ∇ f (x, y) = + (7)
∂x ∂y
patterns in the database, it can determine an appropriate disease
for any type of image and extract proper advices and suggestions. where ∇ f (x, y) is the length of gradient vector, b (x, y) and e (x, y)
It is imperative to achieve high performance and accuracy. There- are a binary image and intensity of the X-ray image respectively. T1
fore, the current work limits the experimental datasets of 5 popular is a threshold. These features are normalized in the forms:
diseases which are evident clinically and radiographically to the e (x, y)
clinicians so that performance of the system will be clearly vali- E (x, y) =   (8)
max e (x, y)
dated [17].
g (x, y)
G (x, y) =   (9)
2.2. Dental extracted features max g (x, y)

For each dental X-ray image, basic image functions, namely: • Patch level feature (Patch): This feature was used to calculate
Local Binary Patterns (LBP) [1], Entropy, edge-value and inten- all gradient vectors of pixels in a patch P, denoted by ı(z) [14].
sity features (EEI) [25], Patch Level Feature (PAT), Red-Green-Blue 
(RGB), and the Gradient (GRA) [14] were applied to extract the fea- Fh (P) = m̃ (z) ı (z) (10)
ture representations, thus obtaining 56 images’ features database. z∈P
These functions achieved good performance in medical applica-
m (z)
tions [30]. In the current work, five extracted features of the dental m̃ (z) =  (11)
images are deliberated, namely:
m(z)2 + εg
z∈P
• Local Patterns Binary feature (LBP): In order to intensify the
difference between areas/clusters in the X-ray dental image, a where m(z) is the gradient amplitude at pixel z, m̃ (z) is the nor-
special case of the texture spectrum model is used. This feature is malization value of m(z) and εg is a small constant. ı(z) is often
known as the LBP feature, which ensures the pixel density order in specified by Hard Binning method as follows.
a given area and is considered to be invariant to any light intensity ⎧
⎨ 1, d (z)
transformation [1].  =i−1
ıi (z) = 2 (12)
⎩
 7 0, otherwise
LBP (xc , yc ) = s (gn − gc ) 2n (1)
• Red-Green-Blue (RGB): The RGB features measure the X-ray
n=0
 image color, which is divided into three matrices based on the
1 x≥0 Red-Green-Blue values [14]. For a 24 bit image, the RGB feature
s(x) = (2) is computed as follows (N is the number of pixels):
0 otherwise
 
where gc is value of the central pixel (xc , yc ) and gn is value of nth hR,G,B [r, g, b] = N ∗ Prob R = r, G = g, B = b (13)
pixel in the window. There is another way to calculate the RGB feature that is isolating
three matrices hR [], hG [] and hB [] with values being specified from
• Entropy, edge-value and intensity (EEI): In order to character- the equivalent color band in the image.
ize the X-ray image structure, the EEI features are extracted to
classify the image into three separated regions, namely: dental • Gradient feature (GRA): In order to differentiate the various tiny
structure, teeth areas, and background [25]. In EEI, entropy is a teeth’s parts such as cementum, enamel, root canal, and the gum;
measure for the randomness level of achieved information within the GRA is used [14]. The following steps calculate the Gradi-
a certain extent and is calculated by equation: ent value: Firstly, we apply Gaussian filter to the X-ray image to
reduce the background noises. After that Difference of Gaussian

L (DoG) filter is applied to calculate gradient of the image accord-
r (x, y) = − p (zi ) log2 p (zi ) (3) ing to x and y axes. Each pixel is the characterized by a gradient
i=1 vector. Lastly, we get normalization form of the gradient vector
and receive a 2D vector for each pixel as follows.
where z is a random variable, p(zi ) is probability of ith pixel, i = 1,2,
. . ., L (L is the number of pixels). The normalization of r(x,y) is  (z) = [sin ˛, cos ˛] (14)
defined as:
where ␣ is direction of the gradient vector. For instance, length and
r (x, y) direction of a pixel are calculated as follows.
R (x, y) =   (4)
max r (x, y) m (x, y) = 2
(L (x + 1, y) − L (x − 1, y)) + (L (x, y + 1) − L (x, y − 1))
2
(15)
 L.H. Son et al. / Biomedical Signal Processing and Control 39 (2018) 64–73 67

Fig. 2. Examples of segmentation using SSFC-FS [44]: (a) Original image; (b) Segmented image.

Fig. 3. Graph representation between one segment and 4 disease patterns [47].

 
 (x, y) = tan−1 (L (x + 1, y) − L (x − 1, y)) / (L (x + 1, y) − L (x − 1, y)) (16) the thresholding method is the threshold value selection for detect-
ing the foremost part of the dental image. The clustering techniques
L (x, y, k) = G (x, y, k) ∗ I (x, y) (17) divide the objects set within the image into clusters, where pixels in
1 x2 +y2 2 2 a group have a specific degree of similarity that is larger than those
G (x, y, k) = √ e−( )/( ) (18)
2 2 in the other groups. However, these clustering based segmenta-
tion methods are sensitive to noises which require improvement in
where I(x,y) is a pixel vector, G(x,y,k) is a Gaussian function of the
these techniques to achieve higher performance. Thus, fuzzy clus-
pixel vector, * is the convolution operation between x and y,  1 is a
tering can be employed to solve such problem, where a data point
threshold.
may be owned by more than one cluster [2]. Then, a membership
matrix that has elements value indicating the membership degrees
2.3. Dental image segmentation
of the data point to different clusters are determined.
Recently, a novel clustering methods called semi-supervised
Image segmentation divides the original image into
fuzzy clustering was introduced [46], where the fuzzy clustering
objects/regions in order to determining the significant objects’
algorithm is gathered with additional information. The additional
location or different areas’ bound in an image. It is performed by
information is used to supervise and to control the clustering
labeling each pixel in the image, thus, pixels with similar features
process in the semi-supervised fuzzy clustering algorithm. Such
will have the same label. Typically, dental X-ray images consist
additional information includes: i) Class labels, where a part of the
of three key regions, namely: i) Dental structural region, which
data is labeled and others are unlabeled; ii) Must-link and cannot-
consists of gums, bone, and other periodontitis structure and
link constraints that requires two elements to be belong to the same
has medium grayscale values; ii) Teeth region, which has high
cluster, while the cannot-link constraint specifies two elements to
grayscale values and; iii) Background region, which has the small-
be in two different clusters, and iii) pre-defined membership matrix
est grayscale values that displays the teeth structure background.
[43]. For image segmentation problems, the semi-supervised fuzzy
Such dental X-ray images structure makes the segmentation more
clustering approaches use the membership matrix as additional
complex compared to the traditional image segmentation.
information.
Dental X-ray images segmentation has significant applications
including dental age estimation [31], forensic identification, auto-
mated dental identification system, teeth numbering, and dental 2.4. Semi-Supervised fuzzy clustering algorithm using interactive
diseases diagnosis. The main image segmentation techniques cate- fuzzy satisficing (SSFC-FS)
gories include i) image processing based segmentation such as the
boundary based methods, thresholding methods, and region-based The proposed DDS modeled the dental X-Ray image segmenta-
methods; and ii) clustering based segmentation methods including tion in the form of semi-supervised fuzzy cluster using Interactive
K-means, and Fuzzy C-means (FCM) [2]. The major disadvantage of Fuzzy Satisficing. It divides the dental X-ray input image into seg-
68 L.H. Son et al. / Biomedical Signal Processing and Control 39 (2018) 64–73

Fig. 4. Graph clustering results.

ments through combining the image’s feature representation with J1 −O1 J2 −O2 J3 −O3
are given by: 1 (J1 ) = , 2 (J2 ) = , and 3 (J3 ) = .
the additional information, represented by a pre-defined member- O1 −O1 O2 −O2 O3 −O3
ship matrix, into the objective function of semi-supervised fuzzy The combination satisficing function is set as follows:
clustering problem. Thus, it can be considered to be a multi-
Y = g1 1 (J1 ) + g2 2 (J2 ) + g3 3 (J3 ) → min (24)
objective optimization problem having the following central parts:
J1 denoting the basic clustering objective, J2 is representing the where, g1 + g2 + g3 = 1 and 0 ≤ g1 , g2 , g3 ≤ 1.
image spatial information, and J3 which is the semi-supervised Thus, the optimization problem is solved with the previ-
clustering additional information. In order to find the global opti- to obtain the optimal solution a(r) =

ouslymentioned constraints

mal solutions, the interactive fuzzy satisficing (IFS) method was (r) (r)
ajk , N (r) = nk , where r is the number of iteration steps
conducted. The SSFC-FS models the segmentation problem as a multi- S×L L
objective optimization problem to be minimized, which can be using the procedure in [46]. An example of segmentation using the
expressed as: SSFC-FS algorithm can be retrieved in [44] (Fig. 2).
Subsequently, a graph representation is used to present the rela-
J = J1 + J2 + J3 → min (19) tion between disease patterns and segments in a visual way. In
 N C addition, the classification task is performed using the affinity prop-
J1 = u2jk
Xk − Vj
2 (20) agation clustering (APC + ) which is applied to determine the most
k=1 j=1 suitable disease patterns corresponding to the segment from i = 1
  to i > S1 as explained in the following section.

N 
C

N 
C
1
l
J2 = um R2 +
jk jk
um
jk
wik (21) 2.5. The proposed DDS methodology
l
k=1 j=1 k=1 j=1 i=1
The database is divided into training and testing sets by 10 runs

N 
C
m cross validation, where the training set is regarded as the dental
J3 = |ujk − u jk |
Xk − Vj
2 (22)
image database. The performances of these methods were mea-
k=1 j=1
sured and averaged over 10 runs. The proposed DDS framework
which satisfy the constraints procedure is as follows. The SSFC-FS is used to divide the dental
X-ray input image into segments obtaining the feature representa-

C

C
tions of a new image in the testing set into S* segment [46]. Then,
ujk = 1; ujk < 1; ∀k
new procedures of classification and decision making are demon-
j=1 j=1
strated in the sub-sections below.
= 1, N with ujk ∈ [0, 1] ; ∀k = 1, N , ∀j = 1, C (23). Since, not all segments are valuable for diagnosis; for example,
a segment consists of the background (dark) part of the image did
In those equations, N and C denote the numbers of data points and not reveal any related information about the disease. Thus, the most
clusters respectively. ujk and ujk are the membership degree and valuable S1 segments from S* segments is selected. Dental diseases
the pre-defined membership degree of data point Xi to cluster Vj usually occur in the teeth region as well as the gum, bone and sur-
respectively. The pre-defined membership matrix is randomly set round structure regions [29]. Once, a disease on teeth occurs, there
within [0,1] according to the given constraint. m is the fuzzifier are some abnormal pixels on the X-ray image in comparison with
and usually set as 2.
.
denotes the Euclidean norm. Rik is a fuzzy their neighbor pixels. Thus, the intensity among these pixels is not
distance between data point Xk to cluster Vi . l is the number of homogeneous [6]. This observation is used to select S1 segments
dental features and belongs to [1,4]. wi (i = 1, .., l) is the value of by calculating the weights for each segment using the following
dental features stated previously. Herein, l is set as 4. expression:
The solution is to find the a1jk , a2jk , a3jk values for the three objective ⎛ ⎞
functions respectively. Afterward, the maximum and minimum  
wk = average ⎝ max aij ⎠ , k = 1, A1 (25)
values of each objective function among the three solutions are rep- pixel i ∈ segment k
j=1, S
resented by: O j , Oj , j = 1, 2, 3, where the fuzzy satisficing functions
 L.H. Son et al. / Biomedical Signal Processing and Control 39 (2018) 64–73 69

Fig. 5. Dental Diagnosis System.

Fig. 6. The diagnostic results for dental diseases: (a) horinzontal root fracture at the cervical third; (b) impacted tooth between teeth 11 and 21; (c) Tooth 12: pulp necrosis
with big periapical lesion ad resorption; (d) Tooth 37 with deep distal caries approaching healthy pulp. Alveolar socket after tooth 38 removal.
70 L.H. Son et al. / Biomedical Signal Processing and Control 39 (2018) 64–73

(k) Table 2
where aij ∈ [0, 1] is the similarity between pixels i and j in seg-
Disease decision table and synthetic results.
ment k that determined by SSFC-FS [46]. Afterward, the calculated
Disease Disease Disease
weights are normalized using the following equation:
pattern 1 pattern 2 pattern m
wk
Wk =   , k = 1, S1 (26) Segment 1 d11 d12 d1m
max wk
Segment 2 d21 d22 d2m
k=1, S1

Segment n dn1 dn2 dnm

Accordingly, the S* segments are sorted in the increasing order
Image y1 y2 ym
of weights. The segments S1 are selected from the lowest similarity
degree to the highest or based on a threshold . Then we create a
graph representation for S1 segments. Table 3
The algorithm.

2.5.1. Graph representation Algorithm: Dental Diagnosis System (DDS)

Fig. 3 illustrates a graph consists of nodes and edges. The graph Start
Input the dental X-ray image, a pattern database
representation is constructed using the dental image dataset and Extract the dental features
the segment’s features, where the nodes and edges are: one node Segment the images using semi-supervised fuzzy clustering (SSFC-FS) Select the
as segment, while others as dental patterns and the edge is the segments
relation between the segment and a dental disease pattern. Notice Construct a graph representation using the dental image dataset and the segment’s
features
that the segments are taken from the process above in Section 2.4.
Apply vector quantization to compress the graph
Consequently, the graph representation is used for the purpose of Determine the most appropriate disease patterns corresponding to the segment using
presenting the relation between disease patterns and segments in a the Affinity Propagation Clustering (APC + )
visual way. In the case of too large number of nodes, a vector quan- Generate the disease decision table
tization (VQ) is applied in order to decrease the number of nodes in Apply fuzzy aggregation operators
Obtain the final diagnostic decision
the obtained graph [24]. This progress is done via a representation Stop
vector of a group of nodes that have long distance to the chosen
segment. Such nodes are removed from the graph together with
the edges between these nodes and segment. The VQ is an effec- 2.5.3. Decision making procedure
tive mean to accomplish all of these requirements as it guarantees A disease decision table is employed to find out the final diagnos-
minimal information loss [24]. tic results for the disease of the image for further decision making.
In the previous steps, the ith segment is in the same cluster with dis-
2.5.2. Improved affinity propagation clustering ease patterns 1 and 3. The normalized distance is calculated from
Afterward, the graph is classified in order to know which cluster the ith segment to all disease patterns of the cluster, e.g. 1 and 3, as
the ith segment falls into. Since, the affinity propagation clustering follows.
⎧
(APC) is used for the exchanging messages among the data points
⎪ d−1 ik
themselves defining which patterns in the database belong to the ⎨  k ∈ cluster
dik = d−1 ik (29)
same disease group with the considered segment [12].
⎪
⎩ k ∈ cluster
In the current work, an improvement of the affinity propagation
0 otherwise
clustering (APC) [12,13,47,53] called APC+ that integrates a new
force function between a data point and an exemplar is invoked. where dik is the Euclidean distance between the ith segment and
In order to evaluate the degree that a data point belongs to a can- kth disease pattern of the cluster. Then, the disease decision table
didate exemplar point; two messages, namely: the responsibility and synthetic results are generated in Table 2.
and availability are used [13]. Clusters are formed after a num-
ber of exchanging messages between nodes. In the APC+, a new 
n
Wi ∗ dij
force function between a data point and an exemplar to increase
i=1
the transference rate of messages is integrated as follows. yj = , j = 1, m (30)
⎧ 
m n

⎪
⎪ d − dik Wi ∗ dij
⎪
⎨ if d > dik
2dik j=1 i=1
Fik = (27)
⎪
⎪ where, yi (i = 1, m) is the possibility of acquiring disease ith , where
⎪
⎩ − d − dik if d ≤ dik
2dik Wi (i = 1, N) is the weight of it h segment defined in Eq. (25).
  Lastly, the final disease is determined for the image by maxi-
r(i, k):=q(i, k) − max e(i, k ) + q(i, k ) + q(i, k) ∗ e(i, k) ∗ Fik (28)
k =
/ k mum operator,

where, dik is Euclidean distance between a point i and an exemplar
k; d is the average value of the distance from point i to all exemplars.
In addition, r(i, k), e(i, k) and q(i, k) are the responsibility, availability
and similarity between them, respectively. The clustering results
of Fig. 3 can be visualized in Fig. 4 which demonstrates that the
it h segment is in the same cluster with disease patterns 1 and 3
which means that its disease could be one of those of patterns 1
and 3. Notice that the disease patterns are taken from a knowledge
base in history (a pattern database) that consists of similar cases
occurred in the past. This base is helpful in the diagnosis process of
further similar cases.
ymax = max (y1 , y2 , ...., ym ) (31)
From the preceding methodology the steps of the proposed DDS
system are summarized in the following Table 3.
3. Results and discussions
In the current section, the proposed DDS method is validated
and is compared with relevant algorithms, namely: fuzzy infer-
ence system (FIS) [33], fuzzy k-nearest neighbor (FKNN) [40], prim
spanning tree (GCP), kruskal spanning tree (GCK), and the affin-
ity propagation clustering (APC) [47]. Those algorithms have been
 L.H. Son et al. / Biomedical Signal Processing and Control 39 (2018) 64–73 71

Table 4
The final diagnosis for the testing set in a fold.

Data (Image no.) Disease Pattern Recommended

1 2 3 4 5 Disease

4 0.48 0 0.22 0.3 0 1

8 0.33 0.04 0.23 0.25 0.15 1
11 0.28 0.35 0.08 0.29 0 2
22 0.23 0.3 0.09 0.25 0.13 2
23 0.21 0.32 0 0.24 0.23 2
24 0.2 0.21 0.25 0.22 0.12 3
29 0 0.27 0.34 0.24 0.15 3
32 0.15 0.09 0.27 0.27 0.22 3
39 0 0.22 0.27 0.31 0.2 4
40 0.21 0.06 0.26 0.24 0.23 3
41 0.17 0.15 0.2 0.27 0.21 4
44 0.22 0.21 0 0.26 0.31 5
45 0.08 0.29 0 0.31 0.32 5
47 0.23 0.04 0.15 0.26 0.32 5
58 0.5 0 0.24 0.26 0 1
59 0.41 0 0.18 0.26 0.15 1
64 0.21 0.42 0.11 0.26 0 2
65 0.23 0.19 0.31 0.13 0.14 3
79 0.21 0.11 0.34 0.22 0.12 3
78 0.3 0.17 0.18 0.27 0.08 4
86 0 0.27 0.15 0.24 0.34 5
85 0.23 0.09 0.12 0.21 0.35 5

Table 5 1. It applies hybrid approach which could get rids of uncertain,

Evaluation performance of all methods.
GCP GCK APC FIS FKNN
MSE 1.908 1.908 0.845 0.2098 0.2863
MAE 1.002 1.007 0.805 0.1982 0.2346
Accuracy (%) 58.46 58.46 90.01 89.67 80.05
implemented in Matlab 2014 and executed on a PC SONY VAIO lap-
top with Core i5 processor. The source codes and the experimental
datasets have been made publicity through the link in Appendix. In
order to validate algorithms performance, some metrics are mea-
sured including the mean squared error (MSE), mean absolute error
(MAE) and accuracy as validity indices. The DDS user interface is
illustrated in Fig. 5.
Using this system with the configuration of testing, the final
diagnosis is reported for the testing set in a fold in Table 4. For
each image, e.g. Img4, the values of yi (i = 1, 5) corresponding to
5 diseases are shown as 0.48, 0, 0.22, 0.3 and 0, respectively. By
applying Eq. (12), the maximum operator is matched with disease
pattern number 1, i.e Img4 corresponds to Disease ID = 1 or “root
fracture”. Analogously, the values and diseases for other images are
calculated and are matched with their corresponding pattern in the
testing set.
The detected (diagnosed) diseases in Table 4 are then compared
with the correct ones of the images in order to calculate the mean
square error (MSE), mean absolute error (MAE) and accuracy values
for this fold. Consequently, similar steps are performed to compute
the MSE, MAE and accuracy values for others fold from 2 to 10. The
final results shown in Table 5 are the average results of 10 runs.
Table 5 establishes that the DDS has better performance than
other algorithms. It is obvious that DDS has the smallest values
of MSE and MAE compared to the other methods. Moreover, the
DDS accuracy is 92.74% which is superior to the other methods.
In addition, Fig. 6 illustrates the diagnostic results for the dental
diseases.
After localization into segments, it is clearly observed that an
image will assist clinicians to concentrate on the most potential
ones and give accurate decisions of the disease. Generally, in the
view of advancements and differences compared to the existing
methods, DDS achieves the following advantages:
ambiguous and unstable performance due to duplicate, conflict
DDS and meaningless rules in FIS [33].
0.0804 2. It uses the SSFC-FS [46] which is the best dental image segmen-
0.0804 tation method based on semi-supervised fuzzy clustering. It has
92.74 been proven in the equivalent article to have better clustering
quality than other methods such as Otsu [35], fuzzy c-means
(FCM) [2], semi-supervised entropy regularized fuzzy clustering
(eSFCM) [51], and eSFCM − Otsu [43].
3. The DDS uses a new graph-based clustering algorithm called
APC+ for the classification task. The APC+ improves the methods
in the previous work reported in [47], namely: prim spanning
tree (GCP), kruskal spanning tree (GCK), and affinity propagation
clustering (APC) by integrating a new force function between a
data point and an exemplar to increase the transference rate of
messages. Furthermore, by employing segmentation and deci-
sion making methods, the DDS achieves superior accuracy to the
GCP, GCK and APC which work without those methods and deter-
mine the final disease through nearest neighbor mechanism in
Fuzzy K-Nearest Neighbor (FKNN) [40].
4. The DDS makes uses of a new decision making procedure that
determine the final disease from a group of diseases found from
the segments. It is straightforward and different to FKNN [40]
with many possible diseases being aggregated and synthesized
to extract the final one.
5. The DDS makes uses of little number of parameters, and the
parameters are mostly automatically determined in the program
or by expert’s experience. Thus, this makes the algorithm more
effective in terms of parameter controlling.
6. The DDS is easy to implement and is significant to the dental
diagnosis problem and the interdisciplinary medical-informatics
field.
The preceding results establish that the proposed DDS system
outperforms the other algorithms for diagnosing the five dental
diseases with accuracy of 92.74%. Thus, as a further work, the cur-
rent work can be extended in the following ways: i) increasing
diagnostic results with clinical diagnosis results; ii) validating the
performance of DDS on a very large dataset including noises and
outliers; and iii) accelerating computational speed of the system.
In addition, it is recommending to diagnosis other dental diseases
72 L.H. Son et al. / Biomedical Signal Processing and Control 39 (2018) 64–73
as well as to apply the DDS proposed algorithm with other medical
diagnosis problems.
4. Conclusions
In the current paper, a novel framework called Dental Diagno-
sis System– DDS was proposed. It is a hybrid approach combining
segmentation, classification and decision making methods for the
dental diagnosis problem. The proposed DDS utilized the best
dental image segmentation method, namely SSFC-FS for the seg-
mentation task. An improvement of graph-based clustering called
APC+ was presented in the DDS for the classification task aim-
ing to clarify the most similar diseases from the dental image
database. The synthetic results were then processed by a new
decision-making procedure that uses weights of segments and
their membership degree vectors to determine the final disease.
The experimental results on the real dental case of Hanoi Medi-
cal University, Vietnam including 87 dental images of 5 diseases in
the period 2014–2015 established that the DDS has superior per-
formance to the relevant ones including Fuzzy Inference System
(FIS), Fuzzy K-Nearest Neighbor (FKNN), Prim spanning tree (GCP),
Kruskal spanning tree (GCK), and Affinity Propagation Clustering
(APC). The findings of this research are significant to the dental
diagnosis problem and interdisciplinary medical-informatics field
in terms of autonomous recognition systems for dental diseases
which was manually examined by clinicians.
Conflict of interest
The authors declare that they have no conflict of interest.
Ethics statement
All study protocols were approved by Hanoi Medical University
Ethical Committee and have therefore been performed in accor-
dance with the ethical standards laid down in the 1964 Declaration
of Helsinki and its later amendments.
Acknowledgement
This research is funded by Vietnam National Foundation for
Science and Technology Development (NAFOSTED) under grant
number 102.05-2017.302.
Appendix A.
Matlab source codes of all algorithms and experimental data can
be found at the URL: https://www.mathworks.com/matlabcentral/
fileexchange/54135-dental-disease-diagnosis–ffsc-apd-.
References
[1] T. Ahonen, A. Hadid, M. Pietikainen, Face description with local binary
patterns: application to face recognition Pattern Analysis and Machine
Intelligence, IEEE Trans. 28 (12) (2006) 2037–2041.
[2] J.C. Bezdek, R. Ehrlich, W. Full, FCM: The fuzzy c-means clustering algorithm,
Comput. Geosci. 10 (2) (1984) 191–203.
[3] A. Bouchachia, W. Pedrycz, Data clustering with partial supervision, Data Min.
Knowl. Discov. 12 (1) (2006) 47–78.
[4] A. Bouchachia, C. Vanaret, GT2FC: An online growing interval type-2
self-learning fuzzy classifier, IEEE Trans. Fuzzy Syst. 22 (4) (2014) 999–1018.
[5] D. Cai, X. He, Z. Li, W.Y. Ma, J.R. Wen, Hierarchical clustering of WWW image
search results using visual, textual and link information, Proceedings of the
12th Annual ACM International Conference on Multimedia (2004) 952–959.
[6] D. CARIES, Oral and dental diseases: causes, prevention and treatment
strategies, Burd. Dis. India (2005) 275–278.
[7] S. Chattopadhyay, R.M. Davis, D.D. Menezes, G. Singh, R.U. Acharya, T. Tamura,
Application of Bayesian classifier for the diagnosis of dental pain, J. Med. Syst.
36 (3) (2012) 1425–1439.
[8] C. Chattopadhyay, D.W. Kim, D.S. Gombos, J. Oba, Y. Qin, M.D. Williams, S.P.
Patel, Uveal melanoma: from diagnosis to treatment and the science in
between, Cancer 122 (15) (2016) 2299–2312.
[9] A. Chen, T. Wittman, A. Tartakovsky, A. Bertozzi, Image segmentation through
efficient boundary sampling, in: Sampta’09, 2009 (pp. Special-session).
[10] H. Ding, J. Liu, H. Lu, Hierarchical clustering-based navigation of image search
results, Proceedings of the 16th ACM International Conference on Multimedia
(2008) 741–744.
[11] K. Doi, Current status and future potential of computer-aided diagnosis in
medical imaging, The Br. J. Radiol. 78 (Suppl. 1) (2014) s3–s19.
[12] B.J. Frey, D. Dueck, Clustering by passing messages between data points,
Science 315 (5814) (2007) 972–976.
[13] J. Gao, Z. Ma, Y. Qin, L. Jia, D. Yao, Application of affinity propagation
clustering algorithm in fault diagnosis of metro vehicle auxiliary inverter, in:
2013 International Conference on Electrical and Information Technologies for
Rail Transportation (EITRT2013)-Volume II, Springer Berlin Heidelberg, 2014,
pp. 3–9.
[14] K.H. Ghazali, M.M. Mustafa, A. Hussain, M.E.C. Bandar, G. Kuantan, Feature
Extraction technique using SIFT keypoints descriptors, International
Conference on Electrical and Engineering and Informatics Institut Technology
(2007) 17–19.
[15] S. Gould, T. Gao, D. Koller, Region-based segmentation and object detection,
in: Advances in neural information processing systems, 2009, pp. 655–663.
[16] M. Grabisch, H.T. Nguyen, E.A. Walker, Fundamentals of Uncertainty Calculi
with Applications to Fuzzy Inference, 30, Springer Science & Business Media,
2013.
[17] G. Greenstein, J. Cavallaro, Failed dental implants: diagnosis, removal and
survival of reimplantations, J. Am. Dent. Assoc. 145 (8) (2014) 835–841.
[18] S. Guillaume, Designing fuzzy inference systems from data: an
interpretability-oriented review, IEEE Trans. Fuzzy Syst. 9 (3) (2001) 426–443.
[19] Hanoi Medical University, http://en.hmu.edu.vn/news/default.aspx.
[20] Y. Hu, N. Yu, Z. Li, M. Li, Image search result clustering and re-ranking via
partial grouping, 2007 IEEE International Conference on Multimedia and Expo
(2007) 603–606.
[21] P. Isoardi, R. Ropolo, Measurement of dose?width product in panoramic
dental radiology, The Br. J. Radiol. 76 (902) (2014) 129–131.
[22] N. Kavitha, S. Vijayarathna, S.L. Jothy, C.E. Oon, Y. Chen, J.R. Kanwar, S.1
Sasidharan, MicroRNAs: biogenesis, roles for carcinogenesis and as potential
biomarkers for cancer diagnosis and prognosis, Asian Pac. J. Cancer Prev. 15
(18) (2014) 7489–7497.
[23] M.S. Kavitha, A. Asano, A. Taguchi, T. Kurita, M. Sanada, Diagnosis of
osteoporosis from dental panoramic radiographs using the support vector
machine method in a computer-aided system, BMC Med. Imaging 12 (1)
(2012) 2–11.
[24] H.B. Kekre, T.K. Sarode, Vector quantized codebook optimization using
K-means, Int. J. Comput. Sc. Eng. 1 (2009) 283–290.
[25] Y.H. Lai, P.L. Lin, Effective Segmentation for Dental X-ray Images Using
Texture-based Fuzzy Inference System, Advanced Concepts for Intelligent
Vision Systems, Springer, Berlin Heidelberg, 2008, pp. 936–947.
[26] T. Leung, J. Malik, Contour continuity in region based image segmentation, in:
Computer Vision—ECCV’98, Springer, Berlin Heidelberg, 1998, pp. 544–559.
[27] S. Li, T. Fevens, A. Krzyżak, S. Li, An automatic variational level set
segmentation framework for computer aided dental X-rays analysis in clinical
environments, Comput. Med. Imaging Graph. 30 (2) (2006) 65–74.
[28] L.V. Madoz, M.J. Giuliodori, A.L. Migliorisi, M. Jaureguiberry, R.L. de la Sota,
Endometrial cytology, biopsy, and bacteriology for the diagnosis of subclinical
endometritis in grazing dairy cows, J. Dairy Sci. 97 (1) (2014) 195–201.
[29] M.I. Meurer, L.J. Caffery, N.K. Bradford, A.C. Smith, Accuracy of dental images
for the diagnosis of dental caries and enamel defects in children and
adolescents: a systematic review, J. Telemed. Telecare 21 (8) (2015) 449–458,
http://dx.doi.org/10.1177/1357633X15605225.
[30] H.P. Narkhede, Review of image segmentation techniques, Int. J. Sci. Mod. Eng.
1 (8) (2013) 54–61.
[31] S.J. Nelson, Wheeler’s Dental Anatomy, Physiology and Occlusion, Elsevier
Health Sciences, 2014.
[32] O. Nomir, M. Abdel-Mottaleb, A system for human identification from X-ray
dental radiographs, Pattern Recogn. 38 (8) (2005) 1295–1305.
[33] K.K. Oad, X. DeZhi, P.K. Butt, A fuzzy rule based approach to predict risk level
of heart disease, Glob. J. Comput. Sci. Technol. 14 (3) (2014) 16–22.
[34] K.K. Oad, X. DeZhi, P.K. Butt, A fuzzy rule based approach to predict risk level
of heart disease, Glob. J. Comput. Sci. Technol. 14 (3) (2014) 16–22.
[35] N. Otsu, A threshold selection method from gray-level histograms,
Automatica 11 (285–296) (1975) 23–27.
[36] N.M. Portela, G.D. Cavalcanti, T.I. Ren, Semi-supervised clustering for MR
brain image segmentation, Expert Syst. Appl. 41 (4) (2014) 1492–1497.
[37] J.A. Pretty, G. Maupomé, A closer look at diagnosis in clinical dental practice:
part 1. Reliability, validity, specificity and sensitivity of diagnostic procedures,
J. Can. Dent. Assoc. 70 (4) (2004) 251–256.
[38] A. Priyono, M. Ridwan, A.J. Alias, R.A.O. Rahmat, A. Hassan, M.A.M. Ali,
Generation of fuzzy rules with subtractive clustering, Jurnal Teknologi 43 (1)
(2012) 143–153.
[39] A.E. Rad, M.S.M. Rahim, A. Norouzi, Level set and morphological operation
techniques in application of dental image segmentation, Int. Sch. Sci. Res.
Innov. 8 (4) (2014) 177–180.
[40] E.O. Ramírez Castillo, J. Soria, Hybrid system for cardiac arrhythmia
classification with fuzzy K-Nearest neighbors and neural networks combined
 L.H. Son et al. / Biomedical Signal Processing and Control 39 (2018) 64–73
by a fuzzy inference system, in: Soft Computing for Recognition Based on
Biometrics, Springer, Berlin Heidelberg, 2010, pp. 37–55.
[41] J. Ramírez, J.M. Górriz, D. Salas-Gonzalez, A. Romero, M. López, I. Álvarez, M.
Gómez-Río, Computer-aided diagnosis of Alzheimer’s type dementia
combining support vector machines and discriminant set of features, Inf. Sci.
237 (2013) 59–72.
[42] L.H. Son, N.T. Thong, Intuitionistic fuzzy recommender systems: an effective
tool for medical diagnosis, Knowl. Based Syst. 74 (2015) 133–150.
[43] L.H. Son, T.M. Tuan, A cooperative semi-supervised fuzzy clustering
framework for dental X-ray image segmentation, Expert Syst. Appl. 46 (2016)
380–393.
[44] L.H. Son, T.M. Tuan, Dental segmentation from X-ray images using
semi-supervised fuzzy clustering with spatial constraint, Eng. Appl. Artif.
Intell. 59 (2017) 186–195.
[45] N.T. Thong, L.H. Son, HIFCF: An effective hybrid model between picture fuzzy
clustering and intuitionistic fuzzy recommender systems for medical
diagnosis, Expert Syst. Appl. 42 (7) (2015) 3682–3701.
[46] T.M. Tuan, T.T. Ngan, L.H. Son, A novel semi-Supervised fuzzy clustering
method based on interactive fuzzy satisficing for dental X-Ray image
segmentation, Appl. Intell. 45 (2) (2016) 402–428.
[47] T.M. Tuan, L.H. Son, A novel framework using graph-based clustering for
dental x-ray image search in medical diagnosis, Int. J. Eng. Technol. 8 (6)
(2016) 422–427.
73
[48] R.W. Valachovic, C.W. Douglass, A.B. Reiskin, H.H. Chauncey, B.J. McNeil, The
use of panoramic radiography in the evaluation of asymptomatic adult dental
patients, Oral Surg. Oral Med. Oral Pathol. 61 (3) (1986) 289–296.
[49] E. Yasunori, H. Yukihiro, Y. Makito, M. Sadaaki, On semi-supervised fuzzy
c-means clustering, Proceeding of FUZZ-IEEE 2009 (2009) 1119–1124.
[50] O. Yasunori, D. Mochihashi, T. Matsui, M. Nakano, H. Kameoka, T. Izumitani, K.
Kashino, Bayesian semi-supervised audio event transcription based on
Markov Indian buffet process, in: 2013 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP), 2013, pp. 3163–3167.
[51] X. Yin, T. Shu, Q. Huang, Semi-supervised fuzzy clustering with metric
learning and entropy regularization, Knowl. Based Syst. 35 (2012) 304–311.
[52] C. Yin, J. Xiang, H. Zhang, J. Wang, Z. Yin, J.U. Kim, A new svm method for short
text classification based on semi-supervised learning, 2015 4th International
Conference on Advanced Information Technology and Sensor Application
(AITS) (2015) 100–103 (2015, August).
[53] K. Zhang, X. Gu, An affinity propagation clustering algorithm for mixed
numeric and categorical datasets, Math. Problems Eng. 2014 (2014) 1–8.
[54] J. Zhou, M. Abdel-Mottaleb, A content-based system for human identification
based on bitewing dental X-ray images, Pattern Recogn. 38 (11) (2005)
2132–2142.
[55] G. Zhu, S. Zhang, Q. Zeng, C. Wang, Boundary-based image segmentation
using binary level set method, Opt. Eng. 46 (5) (2007) 050501.