Caries detection using multidimensional projection and neural network

Abstract

Over the past two decades, diagnosis of tooth caries or cavities is considered as one of the emerging research topics. So far, a number of methods are introduced to diagnose the tooth decaying, tooth demineralization and re-mineralization as well. However, the sophistication against the tooth decaying diagnosis arises when the environs are relatively complex. With all this in mind, this paper introduces the caries diagnosing model. Here, the feature extraction is based on Multilinear Principal Component Analysis (MPCA). Further, the classification is done by utilizing renowned classifier named Neural Network (NN). The proposed model is compared with other conventional methods such as the Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Auto Correlation-NN (AC-NN), Gray-Level Co-Occurrence Matrix (GLCM AC-Support Vector Machine (SVM)), and Independent Component Analysis (ICA), and the performance of the approach is analyzed in terms of measures such as Accuracy, Sensitivity, Specificity, Precision, False Positive Rate (FPR), False Negative Rate (FNR), Negative Predictive Value (NPV), False Discovery Rate (FDR), F ${}_{1}$ Score and Mathews correlation coefficient (MCC). Through quantitative analysis, the proposed model proves its efficiency over the conventional methods in detecting caries.

Keywords

Caries MPCA feature extraction classifier tooth decaying

1. Introduction

Image processing is one of the popular technologies that find numerous applications [6, 10, 12, 14, 15]. Among such applications, the medical field is the most useful real-life application, where researchers pay huge interest. The medical image processing mainly adopts various image segmentation algorithms [9], registration and classification algorithms for diagnostic purposes. The processes include image analysis, tracking and segmenting. Advanced methodologies are also adopted to improve the computational efficiency of the image processing applications [11, 13].

The medical image processing applications include tumor detection, segmentation of lung portions [21], investigation on cervical cytological images [22] and much more. The applications are also extended to estimating the growth stages of cells by determining the intracellular patterns, cell size and other parameters that have an impact on cell growth [8]. Techniques on computed tomography (CT) images of the chest can lead to estimate the heartbeat rate [7]. The diagnostic procedures adopt a variety of technologies on the images, associated features, and their contents. For instance, artificial intelligence techniques such as neural network are adopted as a classifier to determine the grades of cancer [5].

Photoacoustic (PA) imaging is emerging as a new modality for non-invasive medical imaging and diagnosis, while tooth decay, often called as caries or cavity, remains as one of the most common oral diseases. Moreover, the health of dental tissue is diagnosed by the PA imaging. Current methods for detecting tooth decay employing the dental explorer and X-ray radiography are subjective. The former researchers mainly discuss extracted tooth decaying diagnosis by ANN (Artificial Neural Network) technology and rule-based system [5]. However, it requires adequate improvement and precision for promising results. Few of the earlier research works are briefly reviewed further.

The main contribution of this paper is to detect the caries on the basis of the advanced image processing algorithms in order to ensure detection accuracy. In the proposed caries detection approach, the acquired dental images are subjected to enhancement followed by required pre-processing. Here, the enhancement algorithm mainly adopts the contrast enhancement to distinguish the white intensity of the teeth. The pre-processing steps dynamically adjust the grey level of the image so that the regions of interest are well focussed. Subsequently, it adopts feature extraction algorithms and a supervised classifier. The proposed feature extraction algorithm describes the images, which are two dimensional, in multidimensional space. Thus, the contents of the images are projected into multi-dimensional format and so the useful information from the images can be required. Then, NN with enhanced learning algorithm will be used to classify the grade of caries.

The paper is organized as follows: Section 2 reviews the literature work; Section 3 describes the framework of the proposed caries detection model. Lastly, Section 4 discusses the results that were obtained, and Section 5 concludes the paper.

2. Literature review

Kang et al. [1] have proposed new methods to measure the tooth demineralization and re-mineralization. It uses Optical Coherence Tomography (OCT) to measure the depth and severity of early lesions. After the OCT, the difference in the depth and integrated reflectivity are detected after demineralization. Moreover, the OCT images of dental decay are exploited to program a scanned laser beam in order to remove decays only. Also, the OCT images are obtained before as well as after removal, and the scans demonstrate the extremely selective removal. As a result, cross-polarization-OCT is best suited for the non-destructive assessment of early demineralization.

Hughes et al. [2] have proposed a laboratory method for the early detection of the tooth decay using ultrasound, which is done by recording B-scans of teeth which exhibit caries lesions with high-frequency ultrasound. This is because of the difference in the relative acoustic impedances of the caries lesion as well as the surrounding healthy enamel. The method also makes use of the signal processing algorithm. The demineralized enamel will produce a moderate reduction coefficient, and it produces a larger difference in the ultrasound.

Li et al. [3] have determined the tooth decay by using the Support Vector Machine (SVM) based diagnosis method. The same can also be done by the rule-based system and Artificial Neural Network (ANN) methods. However, SVM provides better results. Moreover, the SVM is exploited for one step forward by means of dealing with clinical X-ray image on the basis of the tooth decay diagnosis. Finally, the simulation outcome exhibits that the gray-level co-occurrence matrix are better than autocorrelation coefficient.

Sampathkumar et al. [4] proposed a non-contact optical technique for imaging and detection of early-stage dental caries. They have developed a fine-resolution ultra-broadband all-optical photoacoustic imaging (AOPAI) system to image the early stages of tooth decay. AOPAI system provides a non-contact method for early detection of white-spot lesions with a high detection bandwidth. Moreover, AOPAI system presents a non-contact method for early detection of white-spot lesions with a high detection bandwidth, which presents advantages over previously demonstrated ultrasound methods.

Yu et al. [5] have proposed Artificial Neural Network (ANN), to detect the tooth decay. By using the back-propagation (BP) neural network, the X-ray image of patient’s teeth is analyzed. The network attained considerable superior performance in making differential diagnoses between decayed as well as normal teeth with inter-pixel autocorrelation coefficients as its input feature vector. In comparison with the other method, the tooth decay detection accuracy was significantly improved.

Palmier et al. [43] presented the International Caries Detection and Assessment System (ICDAS) and the Post-Radiation Dental Index (PRDI), which is limited viability for the estimation of Radiation-Related Caries (RRC) lesions due to their criteria it does not take into consideration qualitative clinical features of RRC, which are very prevalent in their clinical presentation and patterns of progression.

Zain et al. [44] presented the accuracy of Optical Coherence Tomography (OCT) in order to detect naturally occurring non-cavitated fissure caries (NCFC) in totality as well as at different loci by visually assessing cross-sectional OCT scans (B-scan) with an interpretation criterion. Moreover, the agreement between dimensions of NCFC is evaluated and measured with OCT and polarized light microscopy (PLM).

The PS-OCT technique that used in [1] is a non-invasive method for detecting the flaws. It is a simple method in which the output is shown in a three-dimensional manner. However, the images obtained will be less contrast, and it is not realistic in nature. B-scan is a diagnostic test, which produces a cross-section of the output images is used in [2]. The performance of this method is stable, and it provides accurate output. At the time of process execution, the waves have only limited penetrations, and the resolution of the output image is less. The support vector machine used in [3] has been experimentally demonstrated for its flexibility, and it also provides the unique solution. Nevertheless, the support vector machine suffers from its lack of transparency. Moreover, 3D PA [4] method has shown its advantages over image formation in a direct manner and also provides high detection sensitivity. However, the method fails in terms of accuracy. On the other hand, the neural network [5] method is easy to detect the complex nonlinear relationship.

Figure 1.

Architecture diagram of the proposed methodology.

3. Proposed caries recognition methodology

The schematic diagram of the proposed model of dental caries recognition is briefly illustrated in Fig. 1. The image $I$ is given as the input to the pre-processing stage, which is the initial step of the model. Pre-processing comprises three steps; contrast enhancement, grey thresholding, and active contour. The features are extracted from the resultant pre-processed image $I_{p}$ . In order to extract the features from the image, we have adopted a procedure named multi-layer PCA. Further, NN classifier is used for the classification of caries and non-caries dental images.

3.1 Pre-processing

This is the initial stage, which enhances the input image by following three steps.

3.1.1 Contrast enhancement

The resized gray level input image (it is necessary to increase or decrease the total number of pixels of the image) is given for enhancing the contrast of the image, where input image is $m\times n$ and the resized gray level input image is $m_{1}\times n_{1}$ . This process adjusts the intensity of image [41], by which it improves the visibility of the image by regulating the relative darkness and brightness of the image, which is defined in Eq. (1), where low_in and high_in indicate the contrast limits for the input image and low_out and high_out are the contrast limits for the output image. Thus the current grey level image $I^{G}$ reassigns into new grey level (contrast enhanced) image $I_{\textit{new}}^{G}$

$\displaystyle I_{\textit{new}}^{G}=\Bigg{(}\begin{array}[]{l}((I^{G}\!-\!\text% {low}\_\text{in})/(\text{high}\_\text{in}-\text{low}\_\text{in}))^{\gamma}\\ *\left(\text{high}\_\text{out}\!-\!\text{low}\_\text{out}\right)\end{array}% \Bigg{)}+\text{low}\_\text{out}$ (1)

3.1.2 Gray thresholding

The gray thresholding by Otus’s thresholding [40] process determines the threshold for the image which is used to convert grey pixel to either white or black based on the gray intensity. In other words, if the image intensity $I_{i}$ is lesser than the constant $C^{\prime}$ (i.e. $I_{i}<C^{\prime}$ ), it is switched by black pixel else white pixel.

3.1.3 Active contour [39]

Active contour model also referred as snakes it is exploited to create a contour or mask of the objective, a set of coordinates must be initialized. The formed contour is relocated over the image by the driven forces of the image to the boundaries of the particular objective. This process exploits two kinds of driven forces namely internal energy and external energy. The main aim of the internal energy term is to control the deformations made to the snake, and the use of the external energy term is to control the fitting of the contour onto the image. Hence, a set of coordinates $a(n)$ and $b(n)$ are used to create contour $T(n)$ , which is defined in Eq. (2).

$\displaystyle T(n)=(a(n)b(n));T(n)\in I_{\textit{new}}^{G}(a,b)$ (2)

Algorithm:	Pseudo code of MPCA
Input:	Pre-processed image
Output:	Obtained features $f$
Step 1:	Calculate the mean matrix using Eq. (7)
Step 2:	Center the tensor using Eq. (8)
Step 3:	Unfold $\hat{I}_{t}$ into a matrix. The elements in the mode $m$ unfolding matrix are defined as Eq. (9)
Step 4:	For mode-m, discover the eigenvectors via covariance matrix that defines the matrix $P$ , $C^{(m)}$ $=$ $\hat{P}_{(m)}\hat{P}_{(m)}^{P}$ , and consider $Y_{(m)}=[Y_{1},Y_{2},\ldots,Y_{L_{(m)}}]$ as the chosen eigenvectors, where $L_{(m)}$ refers the largest eigenvectors.
Step 5:	Feature selection can be of two types:
	(i) Using the sample $Z_{i(m)}=\left(\bar{A}_{i}-\bar{M}\right)\times Y_{(m)}^{P}$ for varied modes (ii) Using $Z_{i}=Y_{1}^{P}\left(\bar{A}_{i}-\bar{M}\right)Y_{(2)}$ . The features obtained from two approaches are different; in such a way that, the first one offers $f=L_{(1)}\times I_{p}^{2}+L_{(2)}\times I_{p}^{1}$ features while the second one offers $f=L_{(1)}\times L_{(2)}$ features.
Step 6:	The classification is processed under the obtained feature sets.

In contour, the control points’ normalized index is represented as $n$ , and the coordinates of the contour are referred as $(a,b)$ . The aggregate energy of the deformation model is determined in Eq. (3).

$\displaystyle F^{*}=\int_{0}^{1}\left(F^{\text{int }l}T(n)+F^{\textit{image}}T% (n)\right.\left.+F^{\textit{con}}T(n)\right)dn$ (3)

where the interior energy forces of the curve are represented as $F^{\text{int}}$ , $F^{\textit{con}}$ , denotes the exterior restriction, and the energy of the image is denoted as $F^{\textit{image}}$ . The internal energy is termed as the summation of elastic energy and bending energy, and it is determined in Eq. (4), where $\alpha(n)$ represents the adjustable constant which specifies the continuity, $\beta(n)$ represents the adjustable constant that specifies contour curving.

$\displaystyle F^{\text{int}l}=F^{\textit{elastic}}+F^{\textit{bend}}=\alpha(n)% \left|\frac{du}{dn}\right|^{2}+\beta(n)\left|\frac{d^{2}u}{dn^{2}}\right|^{2}$ (4) $\displaystyle F^{\textit{elastic}}=\alpha(T(n)-T(n-1))^{2}dn\;\mathbf{u}$ (5) $\displaystyle F^{\textit{bend}}=\beta(T(n\!-\!1)\!-\!T(n)\!+\!T(n\!+\!1))^{2}dn$ (6)

Hence the resultant pre-processed image via the active contour model is represented as $I_{p}$ . Figure 2 illustrates the input images and its corresponding resultant pre-processed images.

Figure 2.

Preprocessing stage (a) input images, (b) images after contrast enhancement, (c) images after grey thresholding, (d) images after active contour.

3.2 Feature extraction via multi-linear PCA

From the pre-processed image, the features are extracted using MPCA [42]. MPCA is a multi-linear extension of PCA. MPCA is employed in the analysis of arrays i.e., a cube or hypercube of numbers, also referred to as ‘data tensor’. MPCA has better generalization ability than PCA in image reconstruction. With this approach, the image is rearranged into a 3D tensor as $I_{t}\in S^{I_{p}^{1}\times I_{p}^{2}\times I_{p}^{3}}$ , wherere $I_{p}^{1}$ presents the height of the image, $I_{p}^{2}$ denotes the width of the image, $I_{p}^{3}$ refers the number of images used. The multi-linear projection is visually illustrated in Fig. 3.

$\displaystyle\bar{M}=\left(1/I_{p}^{3}\right)\sum_{i=1}^{I_{p}^{3}}\bar{A}_{i}$ (7) $\displaystyle\hat{I}_{t}=\left[\bar{A}_{1}\!-\!\bar{M},\bar{A}_{2}\!-\!\bar{M}% ,\ldots,\bar{A}_{I_{p}^{3}}\!-\!\bar{M}\right]$ (8) $\displaystyle\left(\hat{I}_{t}^{(m)}\right)_{(\textit{index})}=b_{i_{1}i_{2}i_% {3}}$ (9)

where index $=$ $i_{m}$ , $\sum\nolimits_{o=3}^{m+1}(i_{o}-1)+\sum\nolimits_{o=m-1}(i_{o}-1)$ .

Figure 3.

Diagrammatic illustration of multi-linear projection.

3.3 Classification

Since the neural networks are established to be prominent for classification in various applications due to its flexibility than any other classifier, in this paper, the chosen features are applied to the neural network for recognizing caries. Here, the back propagation algorithm model is used for classification. With known classifications, the back propagation trains a given feed-forward multilayer neural network for a given set of input patterns. While each entry of the sample set is presented to the network, the network observes its output response to the sample input pattern. The output response is then compared to the known and desired output as well as the error value is computed. On the basis of the error, the connection weights are adjusted. The back propagation algorithm is on the basis of the Widrow-Hoff delta learning rule in which the weight adjustment is performed through Mean Square Error (MSE) of the output response to the sample input.

The feature set is represented in Eq. (10), where $N$ denotes the number of chosen features.

$\displaystyle E=[f_{1},f_{2},f_{3},f_{4},\ldots,f_{N}]$ (10)

The weight $H$ of the network model is determined by Levenberg-Marquardt (LM) algorithm. The network model is defined in Eq. (11), where the resulting output from $l^{\text{th}}$ node of the layer $j$ is represented as $O_{l}^{(j)}$ . The input to the layer is represented as $E_{i}^{j}$ , $F(\bullet)$ denotes the non-linear function, the total number of input to the $j^{\text{th}}$ layer is represented as $N^{(j)}$ , $B_{i}$ denotes the input bias to the $j^{\text{th}}$ layer, the $x$ and $y$ variables denotes the coefficient of the weight $H$ , which is shown in Eq. (12). The network output $\hat{G}$ is defined in Eq. (13), where $W^{(h)}$ is the weight of the hidden neurons and ${W^{0}}$ refers the weight of the bias.

$\displaystyle O_{l}^{(j)}=F\left[x_{l}^{(j)}B_{j}+\sum_{i=1}^{N^{(j)}}E^{% \prime}_{i}{}^{(j)}y_{il}^{(j)}\right]$ (11) $\displaystyle H=[x;y]$ (12) $\displaystyle\hat{G}=W^{0}+\sum_{i=1}^{N^{(j)}}O_{l}^{(j)}W_{i}^{(h)}H$ (13)

To provide the training to the network, weight $H^{*}$ is optimally nominated by decreasing the objective function, which is given in Eq. (14), where $G$ is the actual output.

$\displaystyle H^{*}=\arg\min[H]\left\|G-\hat{G}\right\|$ (14)

The classifier results in the detection of caries or non-caries dental images.

Figure 4.

Performance measure of the classifier with varying threshold ranges (a) accuracy of NN for test cases 1–3, (b) sensitivity of NN for test cases 1–3, (c) specificity of NN for test cases 1–3.

Figure 5.

Performance measure of the classifier (a) accuracy of NN for test cases 1–3, (b) sensitivity of NN for test cases 1–3, (c) specificity of NN for test cases 1–3.

4. Results and discussion

The experimentation was carried out in MATLAB 2015a. The database was downloaded from [45], which was categorized into three sets (test cases 1–3) randomly for the analysis purpose. Such test case divisions are required to ensure fair experimentation. Each test case consists of 40 caries images. Here, 28 images are employed for training data and the residual 12 images are employed for test data. Further, the performance of the proposed model was investigated in correspondence with ten measures like accuracy, sensitivity, specificity, and precision, FPR, FNR, NPV, FDR, F ${}_{1}$ Score and MCC. The rate of the accuracy is determined by Eq. (15).

$\displaystyle\textit{Accuracy}=\frac{(\textit{TP}+\textit{TN})}{(\textit{TP}+% \textit{FP}+\textit{FN}+\textit{TN})}$ (15)

where, TP represents the True Positive; TN represents the True Negative; FP represents the False Positive; FN represents the False Negative.

4.1 Detection performance

The performance of the proposed method is compared to the conventional methods namely (PCA) [37], LDA [36], ICA [38], AC-NN [5] and GAC-SVM [3] in terms of afore-mentioned measures by analyzing varied training percentages including 25%, 50%, and 75% respectively for each test case. The performance comparison of the proposed method over conventional methods for the test case 1 is tabulated in Table 2. From the table, it is observed that the accuracy of the proposed model for 25% training is 35.5% better than ICA, 50.9% and 60% better than LDA and PCA. For 50% training, the proposed model is 44.4% superior to ICA and LDA. Similarly, for 75% training, the introduced method is 60.7% better than ICA, 55.5% and 77.7% superior to LDA and PCA. It is also observed that the FDR of the proposed method is minimized gradually whereas the other methods have increased value. It is observed that, for 25% training, the proposed approach is 51.1% better than ICA and 58%, 68.6% better from LDA and PCA, for 70% training, the introduced approach is 66%, 74.6% and 83% superior to ICA, LDA, and PCA.

The performance analysis of the proposed model for the test case 2 is tabulated in Table 2. From the table, it is clear that the accuracy of the proposed method for 25% training is 75%, 70.1%, and 51.9% better than ICA, LDA, and PCA. For 75% training, the proposed method is 60.7% and 28.3% superior to ICA and PCA. FDR of the proposed method is decreased gradually, and for 25% training, the proposed method is 38.8%, 78%, and 51.1% better from the conventional methods like ICA, LDA, and PCA and for 50% training, the proposed model is 86.6% and 41.1% better than both LDA and PCA. Table 3 shows the performance analysis of the proposed method over existing approaches for the third database. It is observed that for training percentage, 25%, the accuracy of the proposed method is 35.5% superior to ICA, 70.2% better than both PCA and LDA. For FNR measure, the proposed model is 72.7%, 80.6%, and 73.9% better than ICA, PCA, and LDA for 25% training. For 50% training, the proposed model is 50% better than ICA, 72.7% better from LDA.

Table 1
Performance investigation of proposed model over conventional methods for test case 1

Table 2

Performance investigation of proposed model over conventional methods for test case 2

Table 3

Performance investigation of proposed model over conventional methods for test case 3

	Training percentage
	25%						50%						75%
Methods	PCA	LDA	ICA	AC-NN	GAC-SVM	MPCA	PCA	LDA	ICA	AC-NN	GAC-SVM	MPCA	PCA	LDA	ICA	AC-NN	GAC-SVM	MPCA
Accuracy	0.47	0.47	0.59	0.27	0.57	0.8	0.75	0.95	0.6	0.15	0.4	0.7	0.5	0.9	0.56	1	0.6	0.9
Sensitivity	0.54	0.38	0.56	0.15	0.62	0.88	0.78	0.94	0.56	0.06	0.44	0.78	0.5	0.9	0.6	1	0.6	0.9
Specificity	0	1	0.8	1	0.25	0.25	0.5	1	1	1	0	0	0	0	0	0	0	0
Precision	0.78	1	0.94	1	0.84	0.88	0.93	1	1	1	0.8	0.88	1	1	0.9	1	1	1
FPR	1	0	0.2	0	0.75	0.75	0.5	0	0	0	1	1	0	0	1	0	0	0
FNR	0.46	0.62	0.44	0.85	0.38	0.12	0.22	0.06	0.44	0.94	0.56	0.22	0.5	0.1	0.4	0	0.4	0.1
NPV	0	1	0.8	1	0.25	0.25	0.5	1	1	1	0	0	0	0	0	0	0	0
FDR	0.22	0	0.06	0	0.16	0.12	0.07	0	0	0	0.2	0.13	0	0	0.1	0	0	0
F ${}_{1}$ _score	0.64	0.56	0.7	0.27	0.71	0.88	0.85	0.97	0.71	0.11	0.57	0.82	0.67	0.95	0.72	1	0.75	0.95
MCC	$-$ 0.32	0.28	0.26	0.15	$-$ 0.09	0.13	0.19	0.79	0.33	0.08	$-$ 0.33	$-$ 0.17	0	0	$-$ 0.2	0	0	0

The advantages of MPCA over PCA are that it has the potential to serve the same function for analyzing tensor structure data. Further, it performs the feature extraction by determining the multi-linear projection that captures most of the original tensorial input variation, which also avoids the initial resizing step. Due to the multidimensionality projection, the features are distinguished well in multidimensional and so classification can be precise.

4.2 Effect of contrast thresholding

In this Section, the performance of the classifier is determined by varying the contrast threshold ranges including 0.1–0.9, 0.2–0.8, 0.2–0.7, 0.3–0.7, and 0.3–0.8 respectively in terms of accuracy, sensitivity, and specificity for the first, second, and third test cases, which is illustrated in Fig. 4. From Fig. 4a, it is observed that, for the first test case, the accuracy of the NN classifier using the proposed MPCA is more at the ranges like 0.2–0.7, 0.3–0.7, and 0.3–0.8 respectively, for the second database, at the ranges like 0.1–0.9, 0.2–0.8, 0.3–0.7, and 0.3–0.8, the NN classifier with MPCA has higher accuracy rate, whereas the other methods have poor accuracy. For the third test case, the accuracy of the NN using the proposed model is more precisely increased at the ranges such as 0.1–0.9, 0.2–0.8, 0.2–0.7, and 0.3–0.8. Similarly, the performance of the NN classifier is analyzed in terms of measures like sensitivity and specificity, which are illustrated in Fig. 4b and c. Overall results have shown the efficiency of the developed model, whereas the conventional methods have shown its poor performance.

4.3 Classifier configuration

The performance of the NN classifier is analyzed in terms of measures like accuracy, sensitivity, and specificity of varying hidden neurons for the first, second, and third test cases, which is illustrated in Fig. 5. From Fig. 5a, it is observed that the accuracy of the NN classifier using the proposed MPCA model for the test case 1 with 10 hidden neurons is 77.7% and 60% better than ICA and PCA. For 40 hidden neurons, the proposed model is 65%, 29.4%, and 17.8% better than ICA, LDA, and PCA. Similarly, for 50 hidden neurons, the proposed model is 58.3%, 18.7% and 69.6% superior to ICA, LDA, and PCA respectively. For the second test case, the accuracy of the NN classifier using proposed MPCA model for 10 hidden neurons is 90%, 5.5%, and 68.4% better than ICA, LDA, and PCA, for 20 hidden neurons, the accuracy of NN using MPCA is 75% better from ICA, LDA and PCA, and for 30 hidden neurons, NN using proposed model is 2.9% and 27.2% superior to both the LDA and PCA. For 40 hidden neurons, the accuracy of NN using MPCA is 69.0%, 24%, and 56.9% better from ICA, LDA, and PCA respectively. The same investigation is carried out for both sensitivity and specificity for all the three test cases, which are illustrated in Fig. 5b and c respectively. From the figures, it is observed that the performance of NN using proposed MPCA model is more efficient while comparing other conventional methods.

5. Conclusion

This paper has proposed the caries diagnosing model. The features were selected on the basis of Multilinear Principal Component Analysis (MPCA). Further, the classification was done by utilizing renowned classifier named Neural Network (NN). The proposed model was compared with the other conventional methods, and the performance of the approach was analyzed. The analysis was done by varying the training percentages, which have resultant the promising results. The proposed method has attained better detection accuracy for 25% training in test case 1, and it is 48.8% superior to the conventional methods, and this shows the superiority of the developed model.

References

Kang

Jiao

J.J.

Lee

M.H.

Darling

C.L.

and Fried

, Nondestructive assessment of early tooth demineralization using cross-polarization optical coherence tomography, IEEE Journal of Selected Topics in Quantum Electronics 16(4) (July/August 2010).

Hughes

D.A.

Girkin

J.M.

Poland

Longbottom

and Cochran

, Focused ultrasound for early detection of tooth decay, in: Ultrasonics symposium (IUS), 2009 IEEE International, 01 April 2010.

Kuang

Y.-J.

and Ye

W.-P.

, Clinical X-Ray image based tooth decay diagnosis using SVM, in: Proceedings of the Sixth International Conference on Machine Learning and Cybernetics, Hong Kong, 19–22 August 2007.

Sampathkumar

Hughes

D.A.

Otten

and Longbottom

, All-optical photo acoustic imaging and detection of early-stage dental caries, Ultrasonics Symposium (IUS), 2014 IEEE International, 23 October 2014.

Y.-J.

Wang

J.-M.

Lin

D.-H.

and Ye

W.-P.

, Tooth decay diagnosis using back propagation neural network, in: Proceedings of the Fifth International Conference on Machine Learning and Cybernetics, Dalian, 13–16 August 2006.

Contla

P.X.

Kamada

and Takago

, Visual interactive learning system using image processing about multiple programming languages, in: 2016 IEEE 5th Global Conference on Consumer Electronics, Kyoto, Japan, 2016, pp. 1–5.

Chiang

T.C.

Tung

M.H.

Rick

Kurniadi

Wang

Z.H.

and Jong

G.J.

, The non-contact respiratory monitoring system using thermal image processing, in: 3rd International Conference on Green Technology and Sustainable Development (GTSD), Kaohsiung, Taiwan, 2016, pp. 86–92.

Baicu

Ifrim

Barbu

Frangu

and Caraman

, Stage evaluation of cell growth in yeast culture through image processing, in: 20th International Conference on System Theory, Control and Computing (ICSTCC), Sinaia, Romania, 2016, pp. 704–709.

Dan Burdescu

Stanescu

Brezovan

Spahiu

C.S.

and Slabu

, Algorithms for image processing in graph-based volumetric segmentation, in: 7th International Conference on Information, Intelligence, Systems and Applications (IISA), Chalkidiki, Greece, 2016, pp. 1–6.

10.

Kobata

Tada

and Muromaki

, Development of automatic parcel delivery system using image processing techniques, in: International Symposium on Flexible Automation (ISFA), Cleveland, OH, USA, 2016, pp. 292–295.

11.

Seva

Metku

Kim

K.K.

Kim

Y.B.

and Choi

, Approximate stochastic computing (ASC) for image processing applications, in: International SoC Design Conference (ISOCC), Jeju, Korea, 2016, pp. 31–32.

12.

Iorga

Nane

Dalen

E.V.

and Bertels

, An image processing VLIW architecture for real-time depth detection, in: 28th International Symposium on Computer Architecture and High Performance Computing (SBAC-PAD), Los Angeles, CA, USA, 2016, pp. 158–165.

13.

Ramirez

M. Lopez

et al.,A novel methodology for calculating heart rate using images processing,IEEE Latin America Transactions,14(8)(Aug 2016),3522–3527.

14.

Erlandsson

et al., Quantitative wavelet domain image processing of dynamic PET data, in: International Conference of the IEEE Engineering in Medicine and Biology Society, New York, NY, 2006, pp. 2787–2790.

15.

Higa

Tsuha

Onaga

and Nakamura

, Position tracking of the mouth using image processing, in: International Conference of the IEEE Engineering in Medicine and Biology Society, New York, NY, 2006, pp. 3460–3463.

16.

Shivamurthy

P.M.

Nagabhushan

T.N.

and Basavaraj

, A foreground marker based centroid initialized Geodesic active contours for histopathological image segmentation, in: Second International Conference on Cognitive Computing and Information Processing (CCIP), Mysuru, India, 2016, pp. 1–6.

17.

Abdullah

S.M.

Tischer

Wijewickrema

and Paplinski

, Hierarchical mutual nearest neighbour image segmentation, in: 2016 International Conference on Digital Image Computing: Techniques and Applications (DICTA), Gold Coast, Australia, 2016, pp. 1–8.

18.

Wang

Berendt

Jha

and Mandai

, Automated nuclear segmentation in skin histopathological images using multi-scale radial line scanning, in: IEEE Healthcare Innovation Point-Of-Care Technologies Conference (HI-POCT), Cancun, Mexico, 2016, pp. 175–178.

19.

Wen

Zhou

and Yi

, Hybrid methods of particle swarm optimization and spatial credibilistic clustering with a clustering factor for image segmentation, in: IEEE International Conference on Industrial Engineering and Engineering Management (IEEM), Bali, Indonesia, 2016, pp. 1443–1447.

20.

Restaino

Dalla Mura

Vivone

and Chanussot

, Context-adaptive pansharpening based on image segmentation, IEEE Transactions on Geoscience and Remote Sensing 55(2) (Feb 2017), 753–766.

21.

Soliman

et al., Accurate lungs segmentation on CT chest images by adaptive appearance-guided shape modeling, IEEE Transactions on Medical Imaging 36(1) (Jan 2017), 263–276.

22.

Song

et al., Accurate cervical cell segmentation from overlapping clumps in pap smear images, IEEE Transactions on Medical Imaging 36(1) (Jan 2017), 288–300.

23.

Alsam

Farup

and Rivertz

H.J.

, Iterative sharpening for image contrast enhancement, in: Colour and Visual Computing Symposium (CVCS), Gjovik, 2015, pp. 1–4.

24.

Lee

S.L.

and Tseng

C.C.

, Image sharpening using DFT based matrix fractional order differentiator, in: 2015 15th International Symposium on Communications and Information Technologies (ISCIT), Nara, 2015, pp. 73–76.

25.

Tsubai

Mitoda

Fukuda

and Ueno

, An implementation of image sharpening based on morphological operations for ubiquitous echo, in: 2006 International Conference of the IEEE Engineering in Medicine and Biology Society, New York, NY, 2006, pp. 2742–2745.

26.

Wang

Shi

Atkinson

P.M.

and Wei

, Approximate area-to-point regression kriging for fast hyperspectral image sharpening, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 10(1) (Jan 2017), 286–295.

27.

Zhang

and Alemzadeh

, A dental vision system for accurate 3D tooth modeling, in: 2006 International Conference of the IEEE Engineering in Medicine and Biology Society, New York, NY, 2006, pp. 4799–4802.

28.

Ghorayeb

S.R.

and Valle

, Experimental evaluation of human teeth using noninvasive ultrasound: Echodentography, IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control 49(10) (Oct 2002), 1437–1443.

29.

Mahmoud

A.M.

Mukdadi

O.M.

Ngan

and Crout

, High-resolution ultrasound imaging for jawbone surface, in: 1st Middle East Conference on Biomedical Engineering, Sharjah, 2011, pp. 252–255.

30.

Nuansanong

Kiattisin

and Leelasantitham

, Diagnosis and interpretation of dental X-ray in case of deciduous tooth extraction decision in children using active contour model and J48 tree, in: International Electrical Engineering Congress (iEECON), Chonburi, 2014, pp. 1–4.

31.

Vinh

T.Q.

Thanh

B.M.

and Tai

N.N.

, Dental intraoral system supporting tooth segmentation, in: International Conference on Computing, Management and Telecommunications (ComManTel), Ho Chi Minh City, Vietnam, 2013, pp. 326–329.

32.

Wang

and Li

, Tooth separation from dental model using segmentation field, in: 2016 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Orlando, FL, 2016, pp. 5616–5619.

33.

Lai

Zhou

Sun

and Xie

, Clinical study of Cone Beam CT imaging and 3D reconstruction aided detection accuracy of impacted tooth, in: 2013 IEEE Third International Conference on Information Science and Technology (ICIST), Yangzhou, 2013, pp. 17–21.

34.

Yang

Zhang

Wang

and Rao

, Image registration of rabbit tooth soft tissue based on B-Spline, in: 2011 4th International Congress on Image and Signal Processing, Shanghai, 2011, pp. 1104–1107.

35.

Y.J.

Wang

J.M.

Lin

D.H.

and Ye

W.P.

, Tooth decay diagnosis using back propagation neural network, in: 2006 International Conference on Machine Learning and Cybernetics, Dalian, China, 2006, pp. 3956–3959.

36.

Duin

R.P.W.

and Loog

, Linear dimensionality reduction via a heteroscedastic extension of LDA: the Chernoff criterion, in: IEEE Transactions on Pattern Analysis and Machine Intelligence 26(6) (June 2004), 732–739.

37.

Lazcano

Madroñal

Salvador

Desnos

Pelcat

Guerra

Fabelo

Ortega

Lopez

Callico

G.M.

Juarez

and Sanz

, Porting a PCA-based hyperspectral image dimensionality reduction algorithm for brain cancer detection on a manycore architecture, Journal of Systems Architecture 77 (June 2017), 101–111.

38.

Montefusco-Siegmund

Maldonado

P.E.

and Devia

, Effects of ocular artifact removal through ICA decomposition on EEG phase, in: 2013 6th International IEEE/EMBS Conference on Neural Engineering (NER), San Diego, CA, 2013, pp. 1374–1377.

39.

Bakoš

, in: 5th Slovakian-Hungarian Joint Symposium on Applied Machine Intelligence and Informatics, 25–26 January 2007.

40.

Vala

M.H.J.

and Baxi

, A review on otsu image segmentation algorithm, International Journal of Advanced Research in Computer Engineering and Technology (IJARCET) 2(2) (2013).

41.

Kaur

and Kaur

, Comparison of contrast enhancement techniques for medical image, in: 2016 Conference on Emerging Devices and Smart Systems (ICEDSS), Namakkal, 2016, pp. 155–159.

42.

Plataniotis

K.N.

and Venetsanopoulos

A.N.

, MPCA: Multilinear principal component analysis of tensor objects, IEEE Transactions on Neural Networks 19(1) (2008).

43.

Palmier

N.R.

Ribeiro

A.C.P.

Fonseca

J.M.

Salvajoli

J.V.

Vargas

P.A.

Lopes

M.A.

Brandao

T.B.

and Santos-Silva

A.R.

, Radiation-related caries assessment through the international caries detection and assessment system and the post-radiation dental index, Oral Surgery, Oral Medicine, Oral Pathology and Oral Radiology 124(6) (December 2017), 542–547.

44.

Zain

Zakian

and Chewa

H.P.

, Influence of the loci of non-cavitated fissure caries on its detection with optical coherence tomography, Journal of Dentistry, Available online 31 January 2018.

45.

https://mynotebook.labarchives.com/share/Vahab/MjAuOHw4NTc2Mi8xNi9UcmVlTm9kZS83NzM5OTk2MDZ8NTIuOA==.

Caries detection using multidimensional projection and neural network

Abstract

Keywords

1. Introduction

2. Literature review

3.1 Pre-processing

3.1.1 Contrast enhancement

3.1.3 Active contour [39]

Table 1 Performance investigation of proposed model over conventional methods for test case 1

4.3 Classifier configuration

5. Conclusion

References

Table 1
Performance investigation of proposed model over conventional methods for test case 1