[Downloaded free from http://www.saudiendodj.com on Monday, September 30, 2019, IP: 182.0.231.

197]

Original Article Identification of pulpitis at dental X‑ray periapical

radiography based on edge detection, texture
description and artificial neural networks

Bernard Y. Tumbelaka, Fahmi Oscandar1, Faisal Nur Baihaki, Suhardjo Sitam1,

Mandojo Rukmo2
Departments of Physics, Faculty of Mathematics and Natural Sciences, 1Radiology Dentistry, Faculty of
Dentistry, University of Padjadjaran, Jatinangor, 2Conservative Dentistry, Faculty of Dentistry, University of
Airlangga, Surabaya, Indonesia

Key words:
ABSTRACT
Artificial neural networks, dental X-ray,
edge detection, image entropy, mean Objectives: The aim of the present research was to identify pulpitis
square error, periapical radiography, through periapical radiography by applying edges as basis image
pulpitis, texture description features, the texture description and the artificial neural networks (ANNs).
Materials and Methods: Input image data records of 10 molar and 10 canine
teeth were used. The clinical diagnosis of interest cases were represented
as normal pulp, reversible and irreversible pulpitis, and necrotic pulp. The
following image processing steps were done. First, the data records were
converted digitally and preprocessed as its original image using the Gaussian
Filter to obtain the best smoothed intensity distribution. Second, the local
image differentiation was used to produce edge detector operators, e(x,y)
as the image gradient; ∇f(x,y) providing useful information about the local
intensity variations. Third, these results were analyzed by using the texture
descriptors to obtain digitally the image entropy, H. The fourth step, all were
characterized by the ANNs. Results: The edge detection carried important
information about the object boundaries of pulpal health and pain conditions
in the dental pulp significantly. The image entropy which was identified, the
diagnostic term, was obtained from texture descriptors in the segmentation
regions where the curves of pulp states tent convergence with the normal pulp
line from 4.9014 to 4.6843 decreasing to the reversible and the irreversible
pulpitis line include the nectrotic pulp line from 4.6812 to 4.5926 and then
inputting to the ANNs analysis at the same of mean square error around
Address for correspondence: 0.0003. Conclusions: Referred to these results, the correlation of the
Prof. Bernard Y. Tumbelaka, image entropy and the ANNs analysis could be linearly classified with the
Faculty of Mathematics and Natural
critical point of 4.6827. Finally, it could be concluded that the direct reading
Sciences, University of Padjadjaran,
radiography is better to be digitized in order to provide us the best choice
Jatinangor, Indonesia.
E‑mail: [email protected]
for diagnosis validation.

INTRODUCTION conservative intervention. Currently in Indonesia,

dental decay has often been diagnosed using

I t is a fact that there has been almost no one

to pay attention earlier until the stage in the
decay process is usually too late for preventive and
radiographic techniques.[1] Until now it is impossible
to detect and monitor the stages of dental decay
process of pulpitis by periapical radiography which
Access this article online
has the sensitivity for this type of lesion. Eventually
Quick Response Code: the diseased tissue has been strangling the blood
Website:
vessels of the dental pulp that leads to death. At
www.saudiendodj.com
this point, pulpitis has become the most common
DOI:
cause of toothache followed by death of the pulp
10.4103/1658-5984.138139
and spread of infection through the apical foramina
into the periapical tissues.[2,3] This could be found
Saudi Endodontic Journal • Sep-Dec 2014 • Vol 4 • Issue 3 115
[Downloaded free from http://www.saudiendodj.com on Monday, September 30, 2019, IP: 182.0.231.197]

Tumbelaka, et al.: Identification of pulpitis at dental X‑ray

among all ages especially younger people due to several separate the pulpitis. It could be also separated the
unhygienic factors like smoke, hormonal disturbance in infected line in two regions of interest as the reversible
woman, diabetes, stress, cancer, AIDS, genetic factors, pulpitis and the irreversible pulpitis. The process used
insufficiently nutrients, and drugs.[4] weight input generated that was always trained until
reached the result of learning appropriate data same as
An attempt to find a new method in radiography that identified. For the first training as compared variables
could identify this inflammation of tooth pulp in soft let amount of database. Furthermore, these selected
tissue is worth to try it. The pulp is the inner part of number of database input   would be very influence the
the tooth that consists of blood vessels, nerve endings, analysis results because the identification error would
and connective tissues. The primary objective of pulp be able corrected if only if the weight input database
therapy is to maintain the integrity and health of the were more added.
teeth and their supporting tissues. The indications are
based on the clinical diagnosis of normal pulp, reversible Experimental setup
pulpitis (pulp is capable of healing), symptomatic or These research works were arranged for the
asymptomatic irreversible pulpitis (vital inflamed pulp is preprocessing and processing of the original image to
incapable of healing) and nectrotic pulp or pulpless (pulp obtain the best intensity distribution using Gaussian filter
do not respond to vitality test).[5‑8] It might be interesting and to identify the tooth in the next step by the edge
to try to find these indications that could be obtained detection as the basic image features of the disinfected
from radiographic evidence of periapical radiolucency and infected tooth regions of image segmentation
to diagnose pulpitis. It was impossible to achieve an separately to find the deterministic processing [Figure  1].
accurate diagnosis of the state of the pulp on the basis
When the non‑deterministic cases were obtained, it was
of clinical proof alone. The only accurate method is to
needed to apply the ANNS analysis.
do a histological examination.[9] Therefore, numerous
classifications of pulp disease had only identified by a
Gaussian filter
limited number of clinical diagnostic tests before effective
It was needed to apply a very commonly used filter,
dental treatment is performed. Therefore, our research
which smoothly, reduces noise and gets small details.
interest was aimed to identify pulpitis in the regions of
Equation[1] of a Gaussian function in two dimensions
interest through periapical radiography by applying edges
was the product of two of one‑dimensional Gaussians,
as basis image features, the texture description and the
one in each dimension, Gσ(x, y) as follows:
artificial neural networks (ANNs).
1 −X 2 + Y 2
Gσ(X, Y) = exp [1]
MATERIALS AND METHODS 2 πσ 2 2σ 2

Input image data of 10 molar and 10 canine teeth of

periapical radiography of X‑ray were used. The clinical
diagnosis of interested cases was separated into four
tooth conditions of pulpal diagnose states: normal
pulp, reversible and irreversible pulpitis and necrotic
pulp. The following image processing steps were
done: The first step, the data records were converted
digitally and preprocessed as its original image using
the Gaussian filter [10,11] to obtain the best smoothed
intensity distribution. The second step, the local image
differentiation technique was used that can produce
edge detector operators, e(x,y) as the image gradient;
∇f(x, y) providing useful information about the local
intensity variations. The third step, these results were
analyzed by using the texture descriptors based on
mean and variance analysis that could be used in order
Figure 1: Preprocessing and processing of the original image using
to obtain digitally the image entropy, H.[12‑14] The fourth Gaussian filter and the edge detection as the basic image features
step, all was characterized by the ANNs. [15] We can of the disinfected and infected tooth regions of image segmentation

116 Saudi Endodontic Journal • Sep-Dec 2014 • Vol 4 • Issue 3

[Downloaded free from http://www.saudiendodj.com on Monday, September 30, 2019, IP: 182.0.231.197]

Tumbelaka, et al.: Identification of pulpitis at dental X‑ray

where: x is the distance from the origin in the horizontal Texture descriptors
axis, y is the distance from the origin in the vertical axis, Mostly, an image characteristic depend on its texture
σ is the standard deviation of the Gaussian distribution. used in the region segmentation. There were several
important simple texture descriptors such as the image
It was also needed to use the discrete convolution of histogram pf  (fk), the arithmetic mean, µ in equation [8]
equation [2] as follows: and the variance of the standard deviation square, σ2 in
equation [9] and the image entropy, H in equation [10]
f ∗ g = ∑ f i g x − i [2] as a scalar value representing the entropy of gray scale
i
image, fk having B pixels are given by:
where: f is the image function, g is the Gaussian
function Mean : µ = ∑Bfkpf(fk) [8]
When it was applied to the edge detection, it might k =1
be a derivative of a noisy signal that could cause Variance : σ2 = ∑ B (fk − µ ) pf (fk )
2
[9]
more noise. Therefore, it was needed to apply the
k =1
Gaussian filter before taking a smoothed derivative. The Image Entropy: H ( B )=− ∑ B pf (fk ) ln pf (fk ) [10]
differentiation and convolution of equation [3] both k =1
linear operators where they commuted together as
where: pf(fk) – the image histogram, fk  – is the various
d df dg image intensity levels, 0  ≤  f   ≤  2 B, k – is the pixel
(f ∗ g ) = ∗ g = f ∗ [3]
location, k = 1, 2,…., B, B – gray scale unit, 2 bits, 4
dx dx dx
bits,….
Edge detection
Image edges were the form of local variations of image Usually, entropy was a statistical measure of
intensity that can produce local image differentiation randomness that can be used to characterize the texture
techniques defined as edge detector operators, the image of the input image histogram. The image histogram was
gradient, ∇f(x,y) could be written as equation [4]: calculated within an image region, f.

∇f(x,y) = [fx fy] T, [4] The relation between the average codeword length, L(f),
and the image entropy was very close that could be
where: f x = ∂f/∂x and f y = ∂f/∂xy are the gradient written as equation [11]:
operators, T is the matrix transpose
H  (B) ≤ L(f) ≤ H(B) +1 [11]
This gave useful information about local intensity
variations. The edge detector, e(x,y) was obtained Entropy of the gray scale image contained of thousand
as the magnitude of f (x, y) that can be written as bits of information, representation of intensity. Entropy
equation [5]: converted any class other than logical to unit 8 for the
histogram count calculation so that the pixel values
e(X, Y) = fx 2 ( X 1Y ) + fy 2 ( X 1Y ) [5] were discrete and directly correspond to a bin value.
Using these texture descriptors, it was deterministic
The sum of the absolute values of partial derivatives fx, analysis  that could refuse several descriptor cause
fy could be employed simply as equation [6]: the reversible and the irreversible problems both the
disinfected and the infected tooth regions. Therefore,
e (x, y) = |fx(x, y)| + |fy(x, y)| [6] it was needed to use the ANNs analysis to be the
non‑deterministic (random) analysis.
The local intensity direction could be described by the
direction angle of the phase of f(x,y) as follows in Artificial neural networks
equation [7]: Pulpitis identification was done through ANNs
analysis, where accuracy was the percentage of
φ(x, y) = arctan (fx, fy) [7]
number of true classified compared to all final
Texture descriptor analysis was based on the edge result classified using mean square error  (MSE)
detection features related to the image entropy. defined as follows in equation [12]:
Saudi Endodontic Journal • Sep-Dec 2014 • Vol 4 • Issue 3 117
[Downloaded free from http://www.saudiendodj.com on Monday, September 30, 2019, IP: 182.0.231.197]

Tumbelaka, et al.: Identification of pulpitis at dental X‑ray

1 pulp, reversible and irreversible pulpitis, and necrotic

MSE =  ∑( Ti − O i )2 [12] pulp. There could be several results described as follow.
N i =1
where: N  − number of database, i  =  1,2,3., N, First, the edge detection carried important information
T  −  final data output, O -  true data output. about the object boundaries as normal pulp, reversible
and irreversible pulpitis, and necrotic pulp significantly
RESULTS which could be valuable for the pulp interpretation as
shown in Figures 2a and b.
Results are summarized in Table 1 and Figures 2‑6.
Ten tooth radiographs were used according to the ROI Second, as shown in Figures 3 and 4 and  Table 1,
(region of interest) of 10 cropped samples for normal first by using mean analysis was obtained directly
for normal pulp of region 1 (171.3669) to region
Table 1: The relationship between variance and
8 (171.5720), and normal pulp to pulptis line separated
mean regions for normal pulp, reversible and by reversible of region 8 to region 12 (186.7028),
irreversible pulpitis, necrotic pulp regions irreversible of region 12 to region 17 (185.5005).
Normal Pulpitis Dead Second, variance analysis was also used for normal
tooth tooth pulp of region 1 (62.9592) to region 8 (54.8623), and
Normal Reversible Reversible Irreversible Necrotic normal pulp to pulpitis line separated by reversible
pulp pulp
1 to 8 8 to 11 11 to 12 12 to 17 18 to 20
of region 8 to 12 (55.2579), irreversible of region
Variance 62.9592 54.8623 to 55.2579 to 46.5700 12 to 17 (46.5700). In Figure 5, by using image
to 55.2579 46.5700 to entropy obtained from texture descriptors by mean and
54.8623 38.4427 variance analysis[15] as deterministic and random curve.
Mean 171.3669 171.572 to 186.7028 to 185.5005
to 186.7028 185.5005 to
The curves of normal pulp and pulpitis regions were
171.572 183.4223 figured convergence with normal pulp line from 4.9014
Image 4.9014 to 4.6812 to to 4.6843 decreased to pulpitis line from 4.6812 to
Entropi 4.6843 4.5926
4.5926 with MSE around 0.0003, as shown in Figure 6.

b
Figure 2: Original Images, Gaussian filter images, cropping and edge detection at normal pulp (a) reversible and irreversible pulpitis (b) tooth
regions of image segmentations

118 Saudi Endodontic Journal • Sep-Dec 2014 • Vol 4 • Issue 3

[Downloaded free from http://www.saudiendodj.com on Monday, September 30, 2019, IP: 182.0.231.197]

Tumbelaka, et al.: Identification of pulpitis at dental X‑ray

9DULDQFH 0HDQ




 

 




 
                                       

Figure 3: Curves of variance and mean analysis for normal pulp, reversible, irreversible, necrotic

Figure 4: Correlation between ROI of f and ROI of disinfected and

infected by pulpitis based mean and variance

Figure 5: Curve of image entropy of normal pulp, reversible and

The results could be inputting and expanding our irreversible pulpitis, and necrotic pulp
observation figured convergence at the normal pulp
and pulpitis regions with normal pulp line decreased reversible pulpitis (4.6827 to 4.6565) and the irreversible
to pulpitis line with the critical point by the texture pulpitis (4.6565 to 4.3973).
description and the ANNs analysis at the same of
MSE around 0.0003. Figure 6 gave the nonlinearity DISCUSSION
of the MSE cur ve representing the accuracy of
final output over the true output during the training. Even though the periapical radiographic film was of
Pulpitis identification was done through ANN analysis, high quality, there might be more important information
where accuracy was the percentage of number of that could not be directly detected in regard to the pulp
true classified compared to all final result classified. disease states. The present research involved to reduce
For observation that resulted with the reaching true the diagnosis suspect and to gain accurate diagnosis that
accuracy below 80% gave the identification still wrong. could be used to increase the probability of a successful
Therefore, it was needed to increase the level of true clinical outcome. The biological states of the pulpal
accuracy into limit of 95% that means to have to reach tissues had a tendency to give correlating symptom
the MSE equal to 0.0003. and radiographic diagnosis. Digital image analysis
as part of image technology has been enormously
The current research had owned the advantage developed in dentistry that aimed to interpret periapical
diagnosis to obtain the regions of interest more radiograph.[16,17] In endodontics, the quality of the image
precisely. We could separate the normal pulp line into is very important as it facilitates accurate interpretation
two regions of interest as the pure normal pulp (4.9577 of the morphology of the root and the root canal
to 4.8442) and the impure pulp called the regular space before establishing the treatment and the during
pulpitis (4.8442 to 4.6827). We could also separate the treatment as well as postoperative and long‑term
the pulpitis line into two regions of interest as the evaluation of the outcome of endodontic treatment.[18‑21]

Saudi Endodontic Journal • Sep-Dec 2014 • Vol 4 • Issue 3 119

[Downloaded free from http://www.saudiendodj.com on Monday, September 30, 2019, IP: 182.0.231.197]

Tumbelaka, et al.: Identification of pulpitis at dental X‑ray

with the critical point by the texture description and

the ANNs analysis. It might be interesting to develop
a learning process from number of database changed
for user’s needs with high flexibility and standardization.

Although there were always some exceptions, the

different symptoms usually detect that a patient presents
with the pulpal diseases. Image entropy could be
expanded to multivariate function where the pulp pain in
higher sensitivity degenerated to the state of dead pulpal
tissue. This will be our new and future research goal.

CONCLUSIONS
Figure 6: Diagnose the best training performance in 32 × 32 iterated at
epoch 7 of data testing simulation using by ANNs analysis
The correlation of the image entropy and the ANNs
analysis could be linearly classified with the critical point
The principles for diagnosis of the condition of the of 4.6827. Finally, it could be concluded that the direct
dental pulp are no different from those applied to the reading radiography is better to be digitized in order to
management of other disease conditions. One of the provide us the best choice for diagnosis validation.
essential methods is radiographic examination which is
an important complement to the clinical examination. It ACKNOWLEDGMENTS
provides information about the presence and severity of
caries, deep restorations in relation to the dental pulp as We acknowledge the great collaboration between Faculty of
well as root fractures and periapical changes. Dentistry and Faculty of Mathematics and Natural Sciences
of University of Padjadjaran and Faculty of Dentistry,
University of Airlangga, Indonesia.
Our research started with the dental image quality
problem that was usually degraded because of its
mechanical, optical and electronic noises from its
REFERENCES
medical tools were utilized. A  Gaussian filter smoothes 1. Mulyani D, McIntyre J. Caries inhibitory effect of fluoridated
an image by calculating weighted averages in a filter sugar in a trial in Indonesia. Aust Dent J 2002;47:314‑20.
box. [22] Therefore, we had to perfor m the image 2. Gutmann JL, Baumgartner JC, Gluskin AH, Hartwell GR,
Walton RE. Identify and define all diagnostic terms for
pre‑processing using a Gaussian filter approximated to
Periapical/Periradicular health and disease states. J Endod
its original image as the best intensity distribution that 2009;35:1658‑74.
analyzes its image identification and feature parameters 3. Dos Santos Guimarães C, Sampaio GC, Barros de Almeida E,
linked with the features on object represented in Silva LB, Mourato AP, Alves DF. Influence of Apical Foramen
Cleaning on Periapical Lesion Repair. Rev Fac Odontol Porto
the image. The image features of edge detection Alegre 2011;52:45‑7.
represented and correlated with means and variances 4. Mayo Clinic, Oral health: A window to your overall health,
identifying the state of normal pulp, pulpitis and Mayo Foundation for Medical Education and Research. Available
from: http://www.mayoclinic.org/dental/[Last accessed on
necrotic pulp as a deter ministic solution. These 2013 May 11].
indicated ambiguous results. A nondeteministic solution 5. Abbott P, Yu C. A clinical classification of the status of
was needed using the texture description of the grey the pulp and the root canal system. Aust Dent J
level image analysis statistically by the image entropy. 2007;52 1 Suppl:S17‑31.
6. Levin LG, Law AS, Holland GR, Abbott PV, Roda RS. Identify
The probability results were represented by the image and Define All Diagnostic Terms for Pulpal Health and Disease
entropy, H(B). It has owned the advantage diagnosis States. J Endod 2009;35:1645‑57.
to obtain the regions of interest (ROI) using image 7. Newton CW, Hoen MM, Goodis HE, Johnson BR,
McClanahan SB. Identify and determine the metrics, hierarchy,
entropy H(B) more precisely than diagnoses radiography and predictive value of all the parameters and/or methods used
film directly even though compared with the EPT during endodontic diagnosis. J Endod 2009;35:1635‑44.
standard.[23,24] It could expand our observation at the 8. Endodontics: Colleagues for Excellence Newsletter. Endodontic
Diagnosis. Ame Assoc Endod. Available from: http//:www.aae.
curves of normal pulp, reversible and irreversible
org/colleagues [Last accessed in 2013].
pulpitis, necrotic pulp state regions figured convergence 9. Seltzer S, Bender IB, Ziont ZM. The dynamics of pulp
with normal pulp state line decreases to pulpitis line inflammation: Correlations between diagnostic data and actual

120 Saudi Endodontic Journal • Sep-Dec 2014 • Vol 4 • Issue 3

[Downloaded free from http://www.saudiendodj.com on Monday, September 30, 2019, IP: 182.0.231.197]

Tumbelaka, et al.: Identification of pulpitis at dental X‑ray

histological findings in the pulp. Oral Surg Oral Med Oral 19. Lozano A, Forner L, Llena C. In vitro comparison of root‑canal
Pathol 1963;16:969‑77. measurements with conventional and digital radiology. Int Endod
10. Mathematica, Gaussian Filter function. Available from: http:// J 2002;35:542‑50.
www.reference.wolfram.com/mathematica/ref/GaussianFilter. 20. Jung M, Lommel D, Klimek J. The imaging of root canal
htm. [Last accessed on 2012]. obturation using micro‑CT. Int Endod J 2005;38:617‑26.
11. Fisher R, Perkins S, Walker A, Wolfart E. Spatial Filters: 21. Yoshioka T, Kobayashi C, Suda H, Sasaki T. An observation of
Laplacian of Gaussian. Available from: http://www.homepages. the healing process of periapical lesions by digital subtraction
inf.ed.ac.uk/rbf/HIPR2/log.htm [Last accessed on 2003]. radiography. J Endod 2002;28:589‑91.
12. Conners RW, Harlow CA. A Theoretical comparison of texture 22. Liu Ce, Szeliski R, Kang SB, Zitnick CL, Freeman WT.
algorithms. IEEE Trans Pattern Anal Mach Intell 1980;2:204‑22. Automatic Estimation and Removal of Noise from
13. Pitas I. Digital image processing algorithms. United States: a Single Image. IEEE Trans Pattern Anal Mach Intell
Prentice-Hall, Inc. Upper Saddle River, NJ, USA. 1993 2008;30:299‑314.
14. Gonzalez RC, Woods RE, Eddins SL. Digital image processing 23. McClanahan SB, Johnson JD, Blundell RE Jr. Pulpal and
using Matlab, NJ. Chap 11. United States: Prentice Hall; 2003. periradicular diagnostic terminology, Naval Postgrad Dental
15. Matlab products. Available from: http://www.mathworks.com/ School National Naval Dental Center, Bethesda, Maryland,
matlabcentral/fileexchange/26694‑gray‑level‑run‑length‑matrix/ Clinical Update, Vol. 24, No. 1 January 2002.
content/GLRLM.[Last accessed on 2012 Dec 12]. 24. American Academy of Pediatric Dentistry (AAPD), Guideline
16. Nair MK, Nair UP. Digital and Advanced Imaging in on Pulp Therapy for Primary and Immature Permanent Teeth,
Endodontics: A Review. J Endod 2007;33:1‑6. Reference Manual, IV 34, No 6, 12/13, 2009. p. 222‑9.
17. Abella F, Patel S, Duran‑Sindreu F, Mercadé M, Bueno R,
Roig M. An evaluation of the periapical status of teeth with
necrotic pulps using periapical radiography and cone‑beam How to cite this article: Tumbelaka BY, Oscandar F, Baihaki FN,
computed tomography. Int Endod J 2014;47:387-96. Sitam S, Rukmo M. Identification of pulpitis at dental X-ray periapical
radiography based on edge detection, texture description and artificial
18. Nance R, Tyndall D, Levin LG, Trope M. Identification of root
neural networks. Saudi Endod J 2014;4:115-21.
canals in molars by tuned‑aperture computed tomography. Int
Endod J 2000;33:392‑6. Source of Support: Nil. Conflict of Interest: None declared.

Announcement
Android App
A free application to browse and search the journal’s content is now available for Android based
mobiles and devices. The application provides “Table of Contents” of the latest issues, which
are stored on the device for future offline browsing. Internet connection is required to access the
back issues and search facility. The application is compatible with all the versions of Android. The
application can be downloaded from https://market.android.com/details?id=comm.app.medknow.
For suggestions and comments do write back to us.

Saudi Endodontic Journal • Sep-Dec 2014 • Vol 4 • Issue 3 121