ICSEA 2016 : The Eleventh International Conference on Software Engineering Advances

Detection and Classification of Dental Caries in X-ray Images Using Deep Neural
Networks

Ramzi Ben Ali, Ridha Ejbali and Mourad Zaied

REGIM-Lab.: REsearch Groups in Intelligent Machines, University of Sfax, ENIS,
Sfax, Tunisia
E-mail: {ramzi.benali.tn, ridha ejbali, mourad.zaied}@ieee.org

Abstract—Dental caries, also known as dental cavities, is the Dental radiography is an important element in the oral
most widespread pathology in the world. Up to a very recent health follow up. It comes in addition to the visual
period, almost all individuals had the experience of this examination done by the dentist. The x-ray allows dentist to
pathology at least once in their life. Early detection of dental "see" what is happening inside of the teeth and bones, thanks
caries can help in a sharp decrease in the dental disease rate. to x-rays of low intensity which can cross these structures.
Thanks to the growing accessibility to medical imaging, the The types of radiography most common used by the dentist
clinical applications now have better impact on patient care. are: the retro-alveolar, the bite-wing, the panoramic
Recently, there has been interest in the application of machine radiograph [3].
learning strategies for classification and analysis of image data.
The machine learning is defined as the ability to make an
In this paper, we propose a new method to detect and identify
dental caries using X-ray images as dataset and deep neural
agent learn how to take a decision on the basis of observations
network as technique. This technique is based on stacked sparse [4]. In the biomedical context, the action of this agent is
auto-encoder and a softmax classifier. Those techniques, sparse reflected by additional information to assist the dentists in
auto-encoder and softmax, are used to train a deep neural making his decision. The patient management is found
network. The novelty here is to apply deep neural network to assigned to several steps, either at the level of diagnosis, of
diagnosis of dental caries. This approach was tested on a real the treatment choice, or also in the surgical intervention. In
dataset and has demonstrated a good performance of detection. the framework of this paper, the agent under focus has a role
to classify biomedical images by machine learning with the
Keywords-dental X-ray; classification; Deep Neural intention of discovering clinically pertinent pathology
Networks; Stacked sparse auto-encoder; Softmax. patterns. These classification operations are based on
decision-making tool. However, the inter patients variability
I. INTRODUCTION poses many challenges for the traditional classification
The radiographs are essential to establish a good algorithms. These have for the most part been configured and
diagnosis and identify several problems that are impossible parameterized on small data sets or on a very specific cohort.
to visualize otherwise. During the last decade, the representations learning, a
In orthodontics, dental radiography that is used frequently sub-domain of the machine learning, has experienced a huge
is the panoramic shot which offers a good "overview" of the comeback particularly in the computer vision domain. These
teeth and jaws and provides the essential information for representations algorithms have especially allowed crossing
screening and diagnosis of several conditions and problems a significant step with regards to the objects recognition [7]
which can be detected at an early age. and to speech recognition [8]. In machine learning, the model
Dental caries is an infectious disease. The enamel of the of Artificial Neural Networks (ANN) is a valuable tool.
tooth is the first affected. A cavity forms in the tooth and then Although the ANN, was invented close to sixty years ago, it
the decay spreads in depth. If the cavity is not treated, the hole still remains an area of active research. Recently, with the
expands and decay can affect the dentin (layer under the deep learning, ANN has in fact allowed to dramatic
enamel) [1]. Pain is beginning to be felt, especially with the improvements in many applications fields such as the
hot, cold or sweet things. Decay can affect the pulp of the computer vision. The increasing amount of available data and
tooth. We then speak about a toothache. Finally, a dental the computing power have made it easier to train high
abscess may appear when the bacteria attack the periodontal capacity models such as deep learning. However, the inherent
ligament, the bone or the gum. difficulties involved in training such models, as an example
Cavities are very frequent. More than nine out of ten the local minima, still have an important impact. The deep
people would have had, at least one, tooth decay. In France, learning thus aims to find solutions through adding some
more than a third of 6 year-old children and more than half of regularization or improving the optimization. Unsupervised
12 year-old children have been affected by this infection. In pre-training or dropout are examples of such solutions.
Canada, 57 percent of children aged between 6 and 12 years Our contribution is as follows: we propose a system of
have had at least one tooth decay [2]. detection and classification of dental caries in X-ray images
using deep neural network. This system can be very useful

Copyright (c) IARIA, 2016. ISBN: 978-1-61208-498-5 223

ICSEA 2016 : The Eleventh International Conference on Software Engineering Advances

for dentists to classify dental X-ray images into tooth decay One of the perspectives of the techniques of deep learning
or normal tooth images. A stacked sparse auto-encoder and a is the replacement of work which still is relatively laborious
softmax classifier [8] are used in our deep neural network. by algorithmic models of supervised learning, non-
This paper will be structured as follows: Section 2 supervised (i.e., not requiring specific knowledge of the
presents the recent work. In Section 3, we will describe our problem studied) or by techniques of extraction of
methodology and demonstrate how to train and classify tooth hierarchical characteristics.
images with deep neural networks with a stacked sparse auto- The idea is very simple: the training system consists of a
encoder and a softmax classifier. In Section 4, we will give series of modules, each one represent a processing step. Each
some results of testing experiments. Finally, Section 5 module can be trained, with adjustable parameters similar to
concludes this paper. the weight of the linear classifiers. The system is trained end-
to-end: to each example, all the parameters of all the modules
II. RELATED WORK are adjusted to approximate the output produced by the
Primarily, the detection of dental caries has been a visual system of the desired output. The deep qualifying term comes
process, principally based on visual-tactile examination and from the arrangement of these modules in successive layers
radiographic examination [1]. In the recent literature, several [12].
techniques have been developed for the detection of dental To be able to train the system in this way, it must be
caries. Kositbowornchai et al. [9] developed a neural network known in which direction and how much to adjust each
to detect artificial dental caries using images from a charged parameter of each module. For this, it is necessary to
coupled device (CCD) camera and intra-oral digital calculate a gradient. The calculation of this gradient is done
radiography. The main disadvantage of this method is that the by the method of back-propagation, practised since mid-
evaluation of the system was done using teeth with artificial 1980s. A deep architecture can be viewed as a multilayer
carries, which are completely different from naturally network of simple elements, similar to the linear classifiers,
affected ones. Saravanan et al. [10] developed a new method inter-connected by training weight. This is what is called a
to detect dental caries in its early stage using histogram and neural network multi-layers.
power spectral analysis. In this method, the detection of tooth The advantage of deep architectures stems from their
cavities is done based on the region of concentration of pixels capacity to learn to represent the world in a hierarchical
with regard to the histogram and based on the magnitude manner. As all layers can be trained, no need to build a
values with regard to the spectrum. The main drawback of characteristics extractor by hand. The training will do it [13].
this study is that this method depends only on the intensity of In addition, the first layers extract some simple characteristics
pixels. Berdouses et al. [11] developed a computer-aided and after that the following layers will combine to form more
automated methodology for the detection and classification and more complex concepts.
of occlusal caries from photographic color images. This
method is based on the segmentation of photographic color III. METHODOLOGY
images. In this section, we will describe and motivate how to train
Even though there are many methods for caries detection and classify tooth images with deep neural networks with
in early stage, it is still necessary to develop accurate carries multiple hidden layers. Multiple hidden layers neural
detection method to help dentist. networks can be very useful in solving classification problems
The problem of the classical approach of forms with complex data, such as images. Each layer can learn
recognition is that it is very difficult to build a good features at a different level of abstraction. We will use, in our
characteristics extractor and that it must be readjusted for deep neural, stacked sparse auto-encoders for features
each new application. The deep learning is a class of methods extraction and a softmax layer to classify the teeth images.
whose principles are known since the end of 1980s, but
whose use was really generalized since approximately 2012.

tooth decay

Tooth Images Autoencoder Learned Features Linear Classifier

Normal Tooth

Figure 1. Shematization of our deep neural networks architecture

Copyright (c) IARIA, 2016. ISBN: 978-1-61208-498-5 224

ICSEA 2016 : The Eleventh International Conference on Software Engineering Advances

between decayed teeth and normal teeth. The SMC's

A. Stacked Sparse Auto-Encoder parameter θ was trained to minimize the cost function. The
A stacked auto-encoder is a neural network consisting of output of this classifier, as we can see in Fig. 1, is two classes:
multiple layers of sparse auto-encoders in which the outputs decayed teeth or normal teeth.
of each layer is wired to the inputs of the successive layer As we can see in Fig. 3, our diagram of the stacked
[14]. network is formed by the encoders from three auto-encoders
In our work, we choose to use stacked sparse auto- and the softmax layer.
encoder (SSAE) in our deep neural network. The size of tooth
image is 64-by-64 pixels. The SSAE is trained, in an IV. EXPERIMENTAL RESULTS AND DISCUSSION
unsupervised fashion, in order to extract hidden features. In this section we report the results of testing experiments
We begin by the training of the first auto-encoder without we carried out to evaluate our deep neural network to classify
use of labels. The size of the input vector in the training data dental X-ray images. The data set of images was collected
of the first auto-encoder is 4096 nodes. It will be minimized from many dentists.
into 300 nodes in the first hidden layer. In the second step, The aim of this work is to classify dental X-ray images
we train another auto-encoder using data as the encoding of into decayed or normal teeth images. So, we have used 1/3 of
the inputs data provided by the previous auto-encoder. In this images for training each class and the rest of other images in
step, also, we decrease the size of the hidden representation the dataset were used for the classification test [14].
to 150, so that the encoder in the second auto-encoder learns The size of input images is 64-by-64 pixels which is an
an even smaller representation of the input data. We repeat entry vector having as dimension 4096. The images have
this step according to the number of the desired layers. been adjusted in such way that the value of the pixels is
This method trains the parameters of each layer between 0 and 1. For each image, there are two possible
individually while freezing parameters for the remainder of labels, corresponding to decayed or normal teeth images.
the model. To produce better results, after this phase of
training is complete, fine-tuning using back-propagation can
be used to improve the results by tuning the parameters of all
layers at the same time.
B. Softmax layer
The softmax classifier (SMC) is important in the field of
machine learning because it can map a vector to a probability
of a given output in binary classification. Softmax classifier
is a supervised model which uses a logistic regression defined
as:
Figure 2. Some Dental X-ray images of our data set
1
h  As we can see in Fig. 2, we can view some of the images
1  exp( T x)
after loading the training data.
For this experience, all considered networks have a
hidden layer containing 3 units and a layer of output. The
Where θ represents a vector of weight, and x is a vector of structure of the different auto-encoders, as we can see in Fig.
input values learned by the previous SSAE. 3, is chosen in a consistent manner. Each experience has been
In the softmax function, we suppose that the labels were repeated 20 times, with 500 iterations for the training and 400
binary: y(i) ∈ {0,1}. We used this classifier to distinguish iterations for the final learning.
X1

h1
X2 X1

h2 X1
X3 X1
....

S
h3
....

X4 X1
....

X1
....

X1

Input image Xn 1st hidden layer 2nd hidden layer 3rd hidden layer

Input raw Pixels Stacked Sparse Auto-encoders Softmax Classfier

Figure 3. Shematization of stacked auto-encoder and a softmax classfier is to classify dental X-ray images into tooth decay or normal tooth images

Copyright (c) IARIA, 2016. ISBN: 978-1-61208-498-5 225

ICSEA 2016 : The Eleventh International Conference on Software Engineering Advances

To measure the quality of our classification system, we result, as approved in Table 1. However, the accuracy and
perform the result by the confusion matrix (Table 1). The reliability of our results can be improved using a larger dental
classification test of our deep neural networks shows very database.
good performance with a rate that reaches 97%. The implementation of training strategies on really deep
structures with several hidden layers is a future extension of
TABLE I. QUALITATIVE RESULT OF OUR CLASSIFICATION APPROACH this work.
Accuracy Target Class
97% Tooth Normal Class ACKNOWLEDGMENT
Decay Tooth precision
The authors would like to acknowledge the financial
Tooth Decay 48 % 1% 98%
Output

support of this work by grants from General Direction of

class

Normal Tooth 2% 49% 96.1%

Class recall 96% 98% Scientific Research (DGRST), Tunisia, under the ARUB
program.
In this Table, 48% of all tooth images are correctly
classified as decayed teeth. Similarly, 49% cases are correctly
REFERENCES
classified as normal teeth. 1% of all images are incorrectly
classified as decayed teeth. Similarly, 2% of all data are [1] M. B. Diniz1, J. Rodrigues, and A. Lussi, “Traditional and Novel
Caries Detection Methods”, Contemporary Approach to Dental Caries,
incorrectly classified as normal teeth. Out of teeth decay Eds Ming-Yu Li, Chapter 6, 2012.
predictions, 98% are correct, and for the normal teeth
[2] C Badet and B Richard, “ Dental caries”, EMC-Dentisterie, 'pp. 40-
predictions, 96.1% are correct. For all teeth decay cases, 96% 48, 2004.
are correctly predicted as decayed teeth. For all normal teeth [3] R. Ben Ali, R. Ejbali, and M. Zaied, “ GPU-based Segmentation of
cases, 98% are correctly classified. Overall, 97% of the Dental X-ray Images using Active Contours Without Edges”, 15th
predictions are correct. International Conference on Intelligent Systems Design and
Applications, pp. 505 – 510, 2015.
[4] H. M.. El Damanhoury, K. S. Fakhruddin, and M. A. Awad,
“Effectiveness of teaching International Caries Detection and
Assessment System II and its e-learning program to freshman dental
students on occlusal caries detection,” European journal of dentistry ,
pp. 493–497, 2014.
[5] A. Vashishth, B. Kaushal, and A. Srivastava, “ Caries Detection
Technique for Radiographic and Intra Oral Camera Images,”
International Journal of Soft Computing and Engineering (IJSCE), pp.
2231-2307, Volume-4, Issue-2, 2014.
[6] Y. Bengio, A. Courville, and P. Vincent, “Representation learning : A
review and new perspectives.”, IEEE transactions on pattern analysis
and machine intelligence, pp. 580-587, 2013
[7] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks”, In NIPS, pp. 1097–1105,
2012.
[8] J. Xu, L. Xiang, Q. Liu, H. Gilmore, J. Wu, J. Tang, and A.
Madabhushi,“ Stacked Sparse Autoencoder (SSAE) for Nuclei
Detection on Breast Cancer Histopathology Images,” IEEE Trans Med
ol. 35, issue 1, pp. 119-130, 2016
[9] S. Kositbowornchai,S. Siriteptawee,S. Plermkamon,S. Bureerat, D.
Figure 4. ROC Curve Chetchotsak, “Anartificial neural network for detection of simulated
dental caries, Int.J.Comput.Assist.Radiol.Surg., pp. 91–96, 2006.
Fig. 4 illustrates the performance of our classifier model [10] T Saravanan, MS Raj, K Gopalakrishnan, “Identification of Early
by showing the TPR (True Positive Rate) against the FPR Caries in Human Tooth Using Histogram and Power Spectral
Analysis”, Middle-East Journal of Scientific Research, pp. 871-875,
(False Positive Rate) for different threshold values. 2014.
[11] E. D. Berdouses, G. D. Koutsouri, E.F. Tripoliti, G. K. Matsopoulos,
V. CONCLUSION C. J. Oulis, and D. I. Fotiadis. A computer-aided automated
In this paper, a deep neural network using Stacked Sparse methodology for the detection and classification of occlusal caries from
photographic color images. Comput Biol Med., pp.119–135, 2015
Auto-encoder framework is presented for classification of [12] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R.
dental X-ray images. The aim of this work is to classify dental Salakhutdinov, “Dropout : A simple way to prevent neural networks
X-ray images into decayed or normal teeth images. from overfitting”, The Journal of Machine Learning Research, vol. 15,
In this work, we have used stacked sparse auto-encoder no. 1, pp. 1929–1958, 2014.
containing three hidden layers and a softmax classifier. [13] D. Kumar, A. Wong, and D. A. Clausi, “Lung nodule classification
using deep features in ct images,” Computer and Robot Vision (CRV),
The conclusion we can draw from our experiments is: in 2015 12th Conference on, pp. 133–138, 2015..
comparison with the classic approach to random initialization [14] X. Zhang, H. Dou, T. Ju, Jun Xu, and S. Zhang, “Fusing Heterogeneous
of the network weight, this method promotes a convergence Features from Stacked Sparse Autoencoder for Histopathological
of the network to a better local minimum as well in Image Analysis”, IEEE Journal of Biomedical and Health Informatics,
classification with regression. This method gives a good 2015.

Copyright (c) IARIA, 2016. ISBN: 978-1-61208-498-5 226

ICSEA 2016 : The Eleventh International Conference on Software Engineering Advances

[15] A. E. Rad, I. B. M. Amin, M. S. M. Rahim, and H. Kolivand.

“Computer-Aided Dental Caries Detection System from X-Ray
Images”, In Computational Intelligence in Information Systems
Springer. International Publishing, pp.233-243, 2015.

Copyright (c) IARIA, 2016. ISBN: 978-1-61208-498-5 227