Identify pulpitis at dental X-ray periapical radiography based on edge detection, texture description and artificial neural networks Bernard Y. Tumbelaka1, Fahmi Oscandar2, Faisal Nur Baihaki3, Suhardjo Sitam4, Mandojo Rukmo5 1,3 Faculty of Mathematics and Natural Sciences, University of Padjadjaran, Jatinangor, Indonesia 2,4 Faculty of Dentistry, University of Padjadjaran, Jatinangor, Indonesia 5 Faculty of Dentistry, University of Airlangga, Surabaya, Indonesia Email: [email protected]; [email protected] ABSTRACT Objectives: Our research interest is aimed to identify pulpitis at the dental X-ray periapical radiography by applying edges as basis image features, the texture description and the artificial neural networks (ANNs). Methods: First, we need to convert the radiography data records digitally and to preprocess the input image as its original image where we use the Gaussian Filter to obtain the best intensity distribution. The second step, we use the local image differentiation technique that can produce edge detector operators, e(x,y) as the image gradient; Vf(x,y) providing useful information about the local intensity variations. The third step, we analyze these results by using the texture descriptors to obtain digitally the image entropy, H. The fourth step, we characterize all by the ANNs. Results: First we obtain that the edge detection carries important information about the object boundaries of pulpitis as disinfected and infected significantly which can be valuable for the pulpitis interpretation. Second, the image entropy obtained from texture descriptors in region segmentation and then inputting to the ANNs analysis where the curves of disinfected and infected regions are figured convergence with disinfected line from 4.9014 to 4.6843 decreases to infected line from 4.6812 to 4.5926 at the same of MSE around 0.0003. Conclusions: Refer to these results, we get that the correlation of the image entropy and the ANNs analysis can be linearly classified with the critical point of 4.6827. Finally, we conclude that the direct reading radiography is better to be digitized in order to provide us the best choice for diagnose validation. Keyword: Pulpitis, Dental X-ray, periapical radiography, edge detection, texture description, artificial neural networks (ANNs), image entropy, and mean square error (MSE).

INTRODUCTION The fact, there has been almost no one to pay attentions earlier until the stage in the decay process is usually too late for preventive and conservative intervention. Currently in Indonesia, dental decay is often diagnosed using radiographic techniques. Until now it is impossible to detect and monitor the stages of dental decay process of pulpitis by periapical radiography which has high sensitivity for this type of lesion. Usually pulpitis can be determined visually by periapical radiography. If the tooth appears intrinsically discolored on the exterior, then pulpitis has occurred. In addition to a complete oral examination, dental X-rays may be obtained to help assess the tooth status. Unfortunately, in 42% of discolored teeth, X-rays appear normal. Likewise humans experiencing discomfort due to pulpitis do not always show changes on X-rays. Therefore, pulpitis is not obtainable only by simple check up but it digitally offers information through its infected level over its disinfected tooth based on image density regions that we can find their indication possibility areas of reversible and irreversible of pulpitis.Our research purpose is how to find a new method in radiography that can identify this inflammation of tooth pulp in soft tissue. The pulp is the inner part of the tooth that consists of blood vessels, nerve endings, lymphatics, and connective tissues. The primary objective of pulp therapy is to maintain the integrity and health of the teeth and their supporting tissues. It is a treatment objective to maintain the vitality of the pulp of a tooth affected by caries, traumatic injure, or other causes. The indications are based on the clinical diagnosis of normal pulp (symptom free and normally responsive to vitality testing), reversible pulpitis (pulp is capable of healing), symptomatic or asymptomatic irreversible pulpitis (vital inflamed pulp is incapable of healing), or nectrotic pulp. We try to find these indications that can be obtained from radiographic evidence of radiolucency to diagnose pulpitis showing the periapical area. It is impossible to reliably achieve an accurate diagnosis of the state of the pulp on clinical proof alone; the only 100% accurate method is treated histological. Therefore numerous classifications of pulp disease

1

have only identified by a limited number of clinical diagnostic before effective dental treatment given. Therefore, our research interest is aimed to identify pulpitis in the regions of interest precisely at the dental X-ray periapical radiography by applying edges as basis image features, the texture description and the artificial neural networks (ANNs). Our research has owned the advantage diagnosis to obtain the regions of interest precisely. We can expand our observation at the curves of disinfected and infected regions figured convergence with disinfected line decreases to infected line with the critical point by the texture description and the ANNs analysis. We are here interesting to develop learning process from number of database changed as user needs with high flexibility results.

METHODS First, we need to convert the radiography data records digitally and to preprocess the input image as its original image where we use the Gaussian Filter to obtain the best intensity distribution. The second step, we use the local image differentiation technique that can produce edge detector operators, e(x,y) as the image gradient; Vf(x,y) providing useful information about the local intensity variations. The third step, we analyze these results by using the texture descriptors based on mean and variance analysis can be used in order to obtain digitally the image entropy, H (Conners, 1980; Pitas, 1993; Gonzalez, 2003). The fourth step, we characterize all by the ANNs (Mathworks, 2012). We can separate the pulpitis. We can also separate the infected line in two regions of interest as the reversible pulpitis and the irreversible pulpitis. The process uses weight input generated that is always trained until reached the result of learning appropriate data same as identified. For the first training as compared variables let amount of database. Furthermore, these selected number of database input will be very influence the analysis results because the identification error will be able corrected if only if the weight input database are more added. Gaussian Filter We need to apply a very common used filter to be smoothing that reduces noise and gets small details. The equation of a Gaussian function in 2-D is the product of two of one dimensional Gaussians, one in each dimension, G(x,y) as follows:

(1) where : x is the distance from the origin in the horizontal axis. y is the distance from the origin in the vertical axis.  is the standard deviation of the Gaussian distribution. We also need to use the discrete convolution as follows:

f  g   f ii g xxii

(2)

ii

where :

f is the image function g is the gaussian function When we apply the edge detection we may have derivatived of noisy signal that can cause more noisy. Therefore we need to apply the Gaussian filter before taking smoothed derivative. The differentiation and convolution both linear operators where they commute together as

d  f  g   df  g  f  dg dx dx dx

(3)

Edge Detection Image edges are the form of local variations of image intensity that can produce local image differentiation techniques defined as edge detector operators, the image gradient, ∇f(x,y) can be written as:

2

∇f(x,y) = [fx fy]T ,

(4)

where: fx =∂f /∂x and fy = ∂f /∂y are the gradient operators T is the matrix transpose This gives us useful information about local intensity variations. We obtain the edge detector, e(x,y) is the magnitude of f(x,y) that can be written as: (5) The sum of the absolute values of partial derivatives fx, fy can be employed simple as: e(x,y) = ∣fx(x,y)∣ + ∣fy(x,y)∣

(6)

The local intensity direction can be described by the direction angle of the phase of f(x,y) as follows: φ(x,y) = arctan(fx,fy)

(7)

Texture descriptor analysis based on the edge detection features related to the image entropy. Texture Descriptors: Mostly, an image characteristic depends on its texture used in the region segmentation. There are several important simple texture descriptors such as the image histogram pf(fk), the arithmetic mean, µ and the variance of the standard deviation square, σ2 and the image entropy, H as a scalar value representing the entropy of gray scale image, fk having B pixels are given by: B

Mean Variance

: µ = ∑ fkpf(fk) : σ2 = ∑ (fk - µ)2pf(fk) k =1

Image Entropy

(8)

k =1 B

(9)

B

: H(B) = - ∑ pf(fk) ln pf(fk)

(10)

k =1

where: pf(fk) – the image histogram fk – is the various image intensity levels, 0 ≤ f ≤ 2

B

k – is the pixel location, k = 1, 2, ….., B B – gray scale unit, 2 bits, 4 bits, ……. Usually, entropy is a statistical measure of randomness that can be used to characterize the texture of the input image histogram. The image histogram was calculated within an image region, f. The relation between the average codeword length, L(f) and the image entropy is very close that can be written as H (B) ≤ L(f) ≤ H(B) + 1

(11)

Entropy of gray scale image contains of thousand bits of information, representation of intensity. Entropy converts any class other than logical to unit 8 for the histogram count calculation so that the pixel values are discrete and directly correspond to a bin value. Using these texture descriptors, we have deterministic analysis that can refuse several descriptors cause the reversible and the irreversible problems both the disinfected and the infected tooth regions. Therefore we need to use the ANNs analysis to be the non deterministic (random) analysis.

3

Artificial neural networks Pulpitis identification is done through ANNs analysis, where accuracy is the percentage of number of true classified compared to all final result classified by using mean square error (MSE) defined as follows: N

MSE =

∑ (Ti  Oi)2

(12)

i =1

where: N  number of database, i = 1,2,3, ...... N T  final data output,

O  true data output. Experimental setup We arrrange our research works for the preprocessing and processing of the original image to obtain the best intensity distribution using Gaussian filter and to identify the tooth in the next step by the edge detection as the basic image features of the disinfected and infected tooth regions of image segmentation separately to find the deterministic processing. When we obtained the non deterministic cases, so we need to apply the ANNS analysis.

Preprocessing

Processing

Figure 1. Preprocessing and processing of the original image using Gaussian filter and the edge detection as the basic image features of the disinfected and infected tooth regions of image segmentation.

RESULTS We obtain several results written as follow: First we obtain that edge detection carries important information about the object boundaries as disinfected and infected by pulpitis significantly which can be valuable for the pulpitis interpretation as shown in Figure 2.

4

Disinfected tooth radiographs

2. (a)

Infected tooth radiographs

2. (b)

Figure 2. Original Images, Gaussian filter images, cropping and edge detection at 2. (a). Disinfected and 2. (b). Infected tooth regions of image segmentations. Variance 65

Mean 190,00

60 55

185,00

50

180,00

45 175,00

40 35

170,00 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Figure 3. Curves of variance and mean analysis for disinfected and infected of pulpitis.

5

Table 1. The relation between variance and mean regions for disinfected and infected pulpitis regions. Disinfected

Infected

disinfected

reversible

reversible

irreversible

Infected/necrotic

1 to 8

8 to 11

11 to 12

12 to 17

18 to 20

Variance

62.9592 to 54.8623

54.8623 to 55.2579

55.2579 to 46.5700

46.5700 to 38.4427

Mean

171.3669 to 171.572

171.572 to 186.7028

186.7028 to 185.5005

185.5005 to 183.4223

Image Entropi

4.9014 to 4.6843

4.6812 to 4.5926

Figure 4. Correlation between ROI of f and ROI of disinfected and infected by pulpitis based mean and variance

Figure 5. Curves of image entropy H(f) of disinfected and infected by pulpitis

Second, as shown in Figure 3 and Table 1, by using mean analysis we obtained directly for disinfected of region 1 (171.3669) to region 8 (171.5720), and disinfected to infected line separated by reversible of region 8 to region 12 (186.7028), irreversible of region 12 to region 17 (185.5005). Variance analysis was also used for disinfected of region 1 (62.9592) to region 8 (54.8623), and disinfected to infected line separated by reversible of region 8 to 12 (55.2579), irreversible of region 12 to 17 (46.5700). By using Table 1 and Figure 4, we can correlated the results with image entropy in Figure 5. Image entropy obtained from texture descriptors by mean and variance analysis (Mathworks, 2012) as deterministic curve and random curve where the curves of disinfected and infected regions are figured convergence with disinfected line from 4.9014 to 4.6843 decreases to infected line from 4.6812 to 4.5926 with MSE around 0.0003. The results can be inputting and expanding our observation figured convergence at the disinfected and infected regions with disinfected line decreases to infected line with the critical point by the texture description and the ANNs analysis at the same of MSE around 0.0003. Figure 6 gives us the nonlinearity of the MSE curve representing the accuracy of final output over the true output during the training. pulpitis identification is done through ANN analysis, where accuracy is the percentage of number of true classified compared to all final result classified. For observation resulted with the reaching true accuracy below 80% gives the identification still wrong. Therefore we needs to increase level of true accuracy into limit of 95% that means to have to reach the MSE equal to 0.0003. Our research has owned the advantage diagnosis to obtain the regions of interest more precisely. We can separate the disinfected line in two regions of interest as the pure disinfected (4.9577 to 4.8442) and the impure disinfected called the regular pulpitis (4.8442 to 4.6827). We can also separate the infected line in two regions of interest as the reversible pulpitis (4.6827 to 4.6565) and the irreversible pulpitis (4.6565 to 4.3973).

6

Figure 6. Diagnose the best training performance in 32x32 iterated at epoch 7 of data testing simulation using by ANNs analysis.

CONCLUSIONS Refer to these results, we get that the correlation of the image entropy and the ANNs analysis can be linearly classified with the critical point of 4.6827. We have found the same regions of interest both for the molar and canine tooth but the decay effect of pulpitis is dominant to the molar tooth. Finally, we conclude that the direct reading radiography is better to be digitized in order to provide us the best choice for diagnose validation.

ACKNOWLEDGEMNT I would like to thank especially to IADMFR and 19th ICDMFR 2013 Organizing Committees in Norway that I have received a Travel Grant from them. Without it I would not be able to joint this event. Also to the collaboration researches with Faculty of Dentistry and Faculty of Mathematics and Natural Sciences, University of Padjadjaran and Faculty of Dentistry, University of Airlangga, Indonesia.

REFERENCES 1. R.W. Conners and C.A.Harlow. 1980. A Theoretical Comparison of Texture Algorithms, IEEE Transactions on Pattern Analysis and Machine Intelligence, 2: 204-222. 2. R.C. Gonzalez, R.E. Woods, S.L. Eddins. 2003. Digital Image Processing Using MATLAB, New Jersey, Prentice Hall, Chapter 11. 3. Pitas, Ioannis, Digital Image Processing Algorithms, Prentice Hall, 1993. 4. http://www.mathworks.com/matlabcentral/fileexchange/26694-gray-level-run-length-matrix/content/GLRLM, Matlab products, accessed 21 December 2012.

7