IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 19, NO. 2, MARCH 2015
571
Discrimination of Tooth Layers and Dental Restorative Materials Using Cutting Sounds Vahid Zakeri, Member, IEEE, Siamak Arzanpour, and Babak Chehroudi
Abstract—Dental restoration begins with removing carries and affected tissues with air-turbine rotary cutting handpieces, and later restoring the lost tissues with appropriate restorative materials to retain the functionality. Most restoration materials eventually fail as they age and need to be replaced. One of the difficulties in replacing failing restorations is discerning the boundary of restorative materials, which causes inadvertent removal of healthy tooth layers. Developing an objective and sensor-based method is a promising approach to monitor dental restorative operations and to prevent excessive tooth losses. This paper has analyzed cutting sounds of an air-turbine handpiece to discriminate between tooth layers and two commonly used restorative materials, amalgam and composite. Support vector machines were employed for classification, and the averaged short-time Fourier transform coefficients were selected as the features. The classifier performance was evaluated from different aspects such as the number of features, feature scaling methods, classification schemes, and utilized kernels. The total classification accuracies were 89% and 92% for cases included composite and amalgam materials, respectively. The obtained results indicated the feasibility and effectiveness of the proposed method.
Fig. 1. Two commonly used dental restorations; (a) amalgam; (b) composite; (c) fractured (broken) composite which should be replaced.
Index Terms—Audio monitoring, audio signal processing, dental restoration, support vector machine (SVM).
I. INTRODUCTION OOTH is an inhomogeneous structure which is composed of different layers including enamel, dentin, pulp, and cementum. Tooth decay is one of the most prevalent disorders, and dentists need to remove infected parts to restore tooth functionality. Dental restoration is a process that begins with removing carries and affected tissues to retain the functionality of tooth layers. Air-turbine dental handpieces are high-speed rotary cutting tools that are widely used by dentists during this operation. The next stage in the process is filling the cavity with appropriate restorative materials. There are different materials such as amalgam, composite, glass ionomer, ceramics, and gold that are
T
Manuscript received July 22, 2013; revised December 16, 2013 and March 27, 2014; accepted April 4, 2014. Date of publication April 15, 2014; date of current version March 2, 2015. This work was supported by funds provided through the discovery research grant of Natural Sciences and Engineering Research Council of Canada, and Sable Industries, Inc. V. Zakeri and S. Arzanpour are with the School of Mechatronic Systems Engineering, Simon Fraser University, Surrey, BC V3T 0A3, Canada (e-mail: [email protected]; [email protected]). B. Chehroudi is with the Faculty of Dentistry, University of British Columbia, Vancouver, BC V6T 1Z4, Canada (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/JBHI.2014.2317503
commercially available for dental restorations [1]. Among these materials, amalgam and composite are widely used by dentists (see Fig. 1). The main advantage of composite over amalgam is its aesthetical feature, because it can be produced in a variety of tooth-matching shades, which makes its appearance very similar to a real tooth. Although current dental restoration materials are very durable, external mechanical forces of clenching and grinding, as well as chronic exposure to acids and solvents in the mouth may result in fatigue, cracks and ultimately failure [see Fig. 1(c)]. Moreover, the performance of dental restorations is also subject to several factors such as restorative materials [2]–[4], the type and position of the tooth [5], [6], the restoration’s shape, size, and number of involved surfaces [7], [8], and the patient’s age [2], [8]. If the old restoration collapses, there is a high potential for developing new decay that requires replacing of the old restoration and removal of all carries. Replacing old failing restorations is one of the most frequent procedures performed in dental practices [9]–[11]. This rate has not been declined despite of all the advancements in dental restoration materials. To conduct dental restorative operations, dentists receive extensive trainings to become experts in interpreting their tactile, auditory, and visual senses as their diagnosis tools [12]. They transfer such sensory information to the actual practice through their perception. This perceptual procedure is highly subjective, and depends on the individual abilities and experience [13]. The main problem with these approaches is that human senses have
2168-2194 © 2014 Crown Copyright
572
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 19, NO. 2, MARCH 2015
limited functionality, and often are insufficient for dentists to rely on. For instance, discerning the boundary of restoration is challenging, in particular with tooth-colored composite restorations. In fact, during the removal of old composite restorations, healthy tooth layers in adjacent areas are often inadvertently removed [13]–[16]. Although the visibility issue is less challenging for amalgam, replacing it still results in tooth losses [11]. Loosing healthy parts of a tooth is a major problem, because it reduces the longevity of the restoration [15]. In addition, tooth is one of the few human body parts that has very limited healing ability, and almost all losses will be permanent. Therefore, developing an objective and sensor-based method for dental restorative operations is a promising approach to overcome the limited functionality of human senses. The main purpose of this approach is to discriminate restorative materials and tooth layers during the cutting and removal procedures. To achieve this goal, this paper studies the feasibility of analyzing cutting sounds of restoration processes as a tool to discriminate tooth layers and restorative materials. The difference among the mechanical properties of tooth layers and restorative materials (e.g., hardness, density, elasticity, etc.) [1], [17] implies that their cutting sounds should carry unique features that might be useful to accurately discriminate different materials. Moreover, the cutting sound is a valuable source of information that is readily and easily available during the cutting procedure. Thus far, no prior works have been found on discrimination of tooth and restorative materials based on the cutting sounds. Perhaps the closest study to this work was conducted by Kocher et al. [17]. They used an ultrasonic dental scaler, and measured its oscillations in contact with a tooth. They successfully validated their approach offline, and recognized enamel, cementum, and calculus tissues. However, no information was found in their report on recognizing the noncontact class (i.e., when the tip of the scaler did not have contact with the tooth). As it will be discussed in the next section, it is necessary in our study to define and recognize such a class. In a broad perspective, our proposed method falls into the field of audio-based monitoring. Audio signals have been studied and used in various applications for monitoring purposes in both real-time and offline settings. Trabelsi et al. [18] proposed a tool-wear monitoring procedure based on the sounds generated in metal cuttings. They followed offline approach to discriminate between sharp, worn, and broken tools. Wan et al. [19] developed an automatic pipeline monitoring system that used sound information of road cutters to avoid accidental breakage. The feasibility of their approach was investigated offline, and it was concluded that further tests would be needed for the on-site environments. In another study, Yadav et al. [20] introduced a real-time condition monitoring system based on audio signals for internal combustion engines. They could effectively identify faults within the engine, and differentiate them from healthy conditions. Audio signals have also been employed in dietary and health monitoring systems. Amft et al. [21] presented an automatic
Fig. 2. Proposed methodology for TCN and TAN classification cases; TCN: tooth/composite/noncontact; TAN: tooth/amalgam/noncontact; STFT: shorttime Fourier transform; SVM: support vector machine; RBF: radial basis function.
procedure to predict food weight based on acoustic recordings of chewing. They conducted the tests offline, and concluded that their approach was feasible for solid foods. In another work, Shin et al. [22] developed an automatic system for real-time monitoring of health conditions using cough sounds. In this paper, we present a novel monitoring system for dental restorative operations based on the cutting sounds that can discriminate tooth layers and restorative materials. The composition of this paper is as follows: a systematic theoretical– experimental methodology is proposed in Section II, Section III presents and elaborates the results, and finally, Section IV provides discussions and conclusions. II. MATERIAL AND METHOD Considering two widely used restorative materials, i.e., amalgam and composite, and the fact that a tooth is not filled with both materials simultaneously, we are facing two cases of classification: 1) Case 1: Classification between tooth, composite, and noncontact classes (TCN), 2) Case 2: Classification between tooth, amalgam, and noncontact classes (TAN). It should be noted that noncontact was considered as a “class,” because the classifier should be able to identify the free-running handpiece, and discriminate the sound from those of the contact classes (tooth or composite/amalgam). To investigate TCN and TAN classification cases, a methodology is proposed which is depicted in Fig. 2. In this methodology, seven different stages were considered as follows. A. Data Collection For data collection, an experienced dentist while maintaining similar conditions to actual cutting procedures conducted several tests in a dental clinic. In vitro tests were performed on cubic samples (1 cm ×1 cm ×1 cm) of amalgam and composite, as well as one intact extracted human third molar. The samples were fixed in a metal chuck (clamp) during the cutting procedure. A “W & H Toplight 898le” air-turbine handpiece and a “330 Diamond” bur were used for cutting, which are among common choices for dentists in restorative procedures.
ZAKERI et al.: DISCRIMINATION OF TOOTH LAYERS AND DENTAL RESTORATIVE MATERIALS USING CUTTING SOUNDS
Fig. 3.
(a) Composite, amalgam and tooth samples; (b) metal chuck (clamp) for holding the cutting samples; (c) microphone and air-turbine handpiece.
Fig. 4.
Sample STFT spectrum (noncontact): (a) before filtering, and (b) after filtering.
A high-frequency microphone (GRAS 40be) was attached to the handpiece to record the cutting sounds (see Fig. 3).This arrangement ensured that the relative position of the handpiece and the microphone was fixed for all tests. The sampling frequency was 48 kHz which was high enough to capture the maximum speed of the handpiece (∼5000 r/s) based on Nyquist–Shannon sampling theorem [23]. A high-speed data acquisition card (LDS Dactron Photon II), and a signal processing software for frequency analysis (RT Pro Photon) were employed to sample and collect generated data on a personal computer. In obtaining the tooth data, the cutting sounds of enamel and dentin layers were recorded. In addition, similar to the actual restorative process, the cutting sounds were recorded with/without water spray (“wet” and “dry”). All cuttings were conducted three times in a parallel plane to the samples’ surface (long axis of the cutting bur was parallel with the long axis of the tooth). The cutting procedure was maintained constant for all samples, and it comprised of three steps. First, the handpiece ran freely for 3 s (noncontact 1) and then the cutting was undertaken for 3 s (contact). In the last step (noncontact 2), the handpiece ran freely again for 2 s. In this cutting procedure, the transition of a noncontact sound to a contact sound and vice versa were considered. B. Preprocessing and Labeling At this stage, the recorded data were filtered using a highpass filter (Butterworth, order 3) with the cut-off frequency at
573
4 kHz to remove the effect of noises and disturbances (e.g., the compressor sound) that happened at lower frequencies. Fig. 4 provides a comparison of a noncontact sample data before [see Fig. 4(a)], and after the filtering [see Fig. 4(b)]. As it can be seen, the filter could efficiently remove the disturbance sound [the first peak in Fig. 4(a)]. The higher peak (around 5000 Hz) indicates the sound pitch of the handpiece when it was running free (noncontact). In each test, the cutting segment was identified manually based on an increase to the signal power and a decrease in the handpiece’s speed. The speed was measured by the method described in [24]. The cutting data then labeled appropriately as tooth, composite, or amalgam. The free-running data of all tests were labeled as noncontact. It should be mentioned that in the labeling process, both dry and wet cutting signals were considered together for each material to include pertinent information. Fig. 5 shows a sample of a recorded sound signal. The cutting period is highlighted by the rectangle, and the rest of data are noncontact. C. Windowing, Feature Extraction, and Feature Scaling After preprocessing and labeling, the next stages of the proposed methodology are windowing, feature extraction, and feature scaling (see Fig. 2).The absolute value of discrete short-time Fourier transform (STFT) coefficients [23] were selected as the features. Hamming windowing with 50% of overlapping was used, because it is a common choice in audio signal processing [25]. The window length (WL) should be chosen small,
574
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 19, NO. 2, MARCH 2015
D. Classification
Fig. 5. Sample of a recorded sound signal; the cutting period is highlighted by the rectangle, and the rest of data are noncontact.
To deal with the classification problem, the machine learning approach was selected. This approach focuses on developing and studying the models that can learn from data [27]. Support vector machine (SVM) is one of the most powerful machine learning algorithms that was chosen in this study. SVM is a popular method for classification that has been employed in various applications such as environmental sound classification [28], [29], cardiac sound classification [30], [31], and speech recognition [32], [33]. SVM is originally a classifier that discriminates data points of two classes by finding a hyperplane that maximizes the separation between the two classes [34]. Each data point belongs to only one class, and is represented by an n-dimensional vector (xi ∈ Rn , i = 1, . . . , m). Each xi is assigned a class-label yi . For a two-class problem, the labels are considered either 1 or −1 (yi ∈ {1, − 1}). The hyperplane takes the format of f (x) = wT x + b
Fig. 6.
Sample STFT spectrum and the averaged coefficients for p = 32.
because we are interested in discrimination between the classes in a small period to be able to prevent the removal of healthy tooth layers. For this study, 2100 data samples were selected in each window, and since the sampling frequency was 48 kHz, the WL was 2100/48 000 ∼ 44 ms. For each window, 2048 Fourier transform coefficients were obtained using the fast Fourier transform algorithm. Since each feature vector had a high dimension (2048 elements), the computation cost could have been increased. To reduce the dimension of each feature vector, a mathematical averaging technique was used. In this technique, 2048 STFT coefficients were divided coefficients in each group). The to p equal-sized groups ( 2048 p data of each group were averaged mathematically to obtain p features. Fig. 6 indicates a sample STFT spectrum and the averaged coefficients for p = 32. The effect of p on the classification accuracy will be explored in the “result” section. Feature scaling is usually utilized to prevent attributes in greater numeric ranges dominate those in smaller numeric ranges [26]. In addition, it can decrease numerical difficulties during the calculation. For this study, two feature scaling methods were investigated: linear scaling and normal scaling. In the linear scaling, all the vector elements were projected linearly to the interval [0, 1], and in the normal scaling, the vector elements were normalized to have a zero mean and unit variance. The effect of these scaling methods on the classification accuracy will be discussed in the “result” section.
(1)
where w and b are the weight vector and the bias (wT denotes the transpose of w). Finding the values of w and b is referred to as the training process, in which SVM uses a selective set of data points that are denoted as training data. After training, another set of data points can be used to evaluate the performance of SVM (testing data). In this study, two different sets of data were used for training and testing. In order to correctly classify the training data, the hyperplane f should be positive for positive data (yi = 1), and negative for negative data (yi = −1). It can be shown that SVM would be formulated as a constrained optimization problem [27] 1 w 2 +C εi 2 i subject to yi wT xi + b ≥ 1 − εi
minw ,ε
(2) εi ≥ 0
(3)
where ε is a slack variable that measures the degree of misclassification, and C is a “hyperparameter” that compromises between classes’ separation and the amount of misclassification. The above-constrained optimization problem can be solved using Lagrange multipliers [27]. One approach for the data points that are not separable in their original space x is to transform them to a new space φ (x) that usually has a higher dimension. In the new space φ (x), the transformed data points will become separable. To reduce the computation cost of a higher dimension space, a technique called kernel method can be employed. The idea of this method is that for some new spaces φ (x), the dot product between the transformed data points (φ(xi ).φ (xj )) is computed efficiently. It can be shown that the solution of (2) and (3) can be rewritten so that it has only dot products of data points (xi .xj ) [27]. Therefore, when data points are transformed to a new space φ (x), the new solution only depends on the kernel function Ker (xi , xj ) = φ(xi ).φ (xj ), and the kernel method is applicable.
ZAKERI et al.: DISCRIMINATION OF TOOTH LAYERS AND DENTAL RESTORATIVE MATERIALS USING CUTTING SOUNDS
575
TABLE I DIFFERENT GROUPS FOR TRAINING AND TESTING
Two commonly used kernels for SVM are radial basis function (RBF) kernel and linear kernel which are described by [27] (4) K (xi , xj ) = exp −γ xi − xj 2 ; γ > 0 and K (xi , xj ) = xi .xj .
(5)
According to (5), in the linear kernel, the original space x and the new space φ (x) are identical. In (4), γ is a hyper-parameter [similar to C in (2)]. To find the optimal values of these hyperparameters, usually a “grid search” approach is employed, in which the values of the hyperparameters are changed, and the optimal values are found based on the cross validation [26]. In a ν-fold cross validation, the training data points are divided to ν equal-sized subsets. Then, ν − 1 subsets are used for training, and the one remained subset is utilized for testing. The entire procedure is performed ν times, and the total classification error is obtained [35]. Those values of C and γ will be selected that result in the minimum error. In this study, the hyperparameters were varied from 2−15 , 2−13 , . . . to 21 , 23 ; and a fivefold cross validation was employed. SVM is basically designed for binary classifications. Therefore, we need to come up with methods to use it for our application, which is a three-class problem. Hsu and Lin compared different methods for multiclass SVM, and concluded that a “one-against-one” (OAO) scheme was more suitable considering practical aspects [36]. In this scheme, β (β − 1) /2 binary classifiers are trained (β is the number of classes), and the class that receives the maximum “votes” is assigned to a test data. III. RESULT A. Selecting the Number of Features, Scaling Method, and SVM Kernel As it was explained in the previous section, SVM first should be trained. Two of the three experiments were used for training, and the third one was employed for testing. Table I indicates different groups of training and testing sets. To apply SVM, the software LIBSVM was used [37]. Considering all classes/groups, the total accuracy of the classifier was obtained for each case (TCN or TAN). As it was mentioned in the “classification” section, the values of hyperparameters [C in (2) and γ in (4)] were found by a fivefold cross validation. The classifier performance was evaluated from different aspects. The number of features (p) was varied as 2, 4, 8, 16, 32, 64, 128, 256, 512, and for each number, the effect of linear/normal scaling methods and linear/RBF kernels on the total accuracy was studied. Figs. 7 and 8 show the results for TCN and TAN cases, respectively.
Fig. 7. Effect of different number of features, scaling methods, and SVM kernels on the total accuracy of TCN case (the number-of-features axis is in logarithmic scale).
Fig. 8. Effect of different number of features, scaling methods, and SVM kernels on the total accuracy of TAN case (the number-of-features axis is in logarithmic scale).
As Fig. 7 indicates for TCN case, increasing the number of features could enhance the classification accuracy. In addition, no particular trends could be observed for different scaling methods and kernels. In fact, depending on the value of p, one selection of scaling and kernel could perform better than others. For example, for p = 16, linear scaling/linear kernel had the maximum accuracy, while for p = 512 normal scaling/RBF kernel was maximum. In general, the maximum accuracy was 89% which was obtained for p = 64 with linear scaling/RBF kernel, and for p = 256, 512, both with normal scaling/RBF kernel. For TAN case as Fig. 8 indicates, by increasing the number of features, the classification accuracy first reached to a maximum, stayed at that maximum for some values of p, and then eventually decreased. Generally, the accuracy order from higher to lower was obtained for normal scaling/linear kernel, linear scaling/linear kernel, linear scaling/RBF kernel, and normal scaling/RBF kernel. The maximum accuracy was 92% which was obtained for p = 32, 64, 128, all with normal scaling/linear kernel. Considering these results and the fact that less number of features causes less computation costs, 64 features with linear scaling/RBF kernel were selected for TCN, whereas 32 features with normal scaling/linear kernel were selected for TAN. Table II displays the hyperparameters’ values of these selections.
576
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 19, NO. 2, MARCH 2015
TABLE II HYPERPARAMETERS’ VALUES FROM FIVE FOLD CROSS VALIDATION
TABLE III CLASSIFICATION ACCURACY OF TCN CASE
TABLE IV CLASSIFICATION ACCURACY OF TAN CASE
All the values of Table II resulted in 98–100% crossvalidation accuracy, and therefore it was not needed to look for other values of C and γ. In this table, no values were obtained for γ in TAN case, because a linear kernel was employed. The hyperparameters of Table II were used to train the SVM classifiers for TCN and TAN cases. Table III (TCN) and Table IV (TAN) indicate the accuracy of these classifiers on the testing data of each group. According to Table III, the worst accuracy for TCN case was 74% which was obtained in classifying of 608 test samples of tooth data in group G2. The best accuracy for this case was for tooth class of group G3 with 98% correct classification out of 567 test samples. For TAN case as Table IV indicates, the worst accuracy was for amalgam of group G1 with 83% correct classification out of 368 test samples. The best accuracy for this case was for both tooth and amalgam classes of group G3 with 97% correct classification out of 567 and 293 test samples, respectively. B. Classification Scheme As it was mentioned in the “classification” section, the OAO scheme was used for SVM. An alternative approach is
Fig. 9.
Hierarchical classification scheme.
considering a hierarchical classification scheme. In this scheme, as indicated by Fig. 9, two levels are considered. In the first level, the classifier decides if the data belong to the contact or noncontact classes. The second level is used for contact data, and based on the classification case, it determines between tooth/composite [case 1 in Fig. 9], or tooth/amalgam [case 2 in Fig. 9]. In order to employ the hierarchical scheme for TCN case, two binary class models for contact/noncontact (Con/N) and tooth/composite (T/C) were obtained. To train these models, similar to OAO scheme, 64 features with linear scaling/RBF kernel were used. In TAN case, two binary class models were obtained for contact/noncontact (Con/N) and tooth/amalgam (T/A). To train these models, 32 features with normal scaling/linear kernel were utilized. In the hierarchical scheme, the values of hyperparameters for both cases were also found by a fivefold cross validation. The developed binary models were employed in the hierarchical scheme (see Fig. 9). Table V (TCN) and Table VI (TAN) show the classification accuracy of this scheme and compare the results with the OAO scheme. As Table V shows, the hierarchical scheme improved the classification accuracy of the tooth class for groups G1 and G2. The accuracy in classification of composite of G2 as well as noncontact classes of G2 and G3 were also increased for the hierarchical scheme. According to Table VI, tooth class of group G2, amalgam class of group G1, and noncontact class of groups G1 and G3, had a better accuracy in the hierarchical scheme. For other classes/groups, the OAO scheme had better performance. The maximum difference between the accuracies of both schemes was for tooth of G1, in which the OAO classification was 17% (96–79%) more accurate than the hierarchical scheme. Considering all groups/classes, the total accuracy in TCN was 89% and 88% for OAO and hierarchical schemes, respectively. In TAN,
ZAKERI et al.: DISCRIMINATION OF TOOTH LAYERS AND DENTAL RESTORATIVE MATERIALS USING CUTTING SOUNDS
577
TABLE V COMPARISON BETWEEN OAO AND HIERARCHICAL CLASSIFICATION SCHEMES FOR TCN CASE
TABLE VI COMPARISON BETWEEN OAO AND HIERARCHICAL CLASSIFICATION SCHEMES FOR TAN CASE
the total accuracy for OAO classification was 92% while for the hierarchical scheme was 91%. C. Temporal Voting Temporal voting is a promising approach to enhance the classification accuracy. As it was mentioned previously in Section II, the recorded sounds were windowed with 50% overlapping. Then, the features were extracted from each window and considered as one sample data. In this approach, the used time-interval was equal to the WL as Fig. 10(a) shows. For this study, the WL was WL = 482100 000 . In the temporal voting approach, the time interval is chosen so that it covers more than one window. As Fig. 10(b) indicates, the time interval is corresponding to five windows (considering 50% of overlapping). In this approach, the class of each sample in the selected time interval is first found separately, e.g., the class of data associated with W1, W2, . . . ,W5. Then, the class that receives the maximum votes will be assigned to the sample associated with W1. If the selected number of windows in the temporal voting approach is nw (with 50% overlapping) and the length of each window is WL, the used time-interval is (nw + 1) WL . 2
(6)
Fig. 10.
Schematic diagram of the temporal voting approach.
To employ temporal voting approach, TCN and TAN classifiers of the OAO scheme were used. Tables VII and VIII compare the effect of temporal voting with different number of windows (nw ) on the classification accuracies of both cases. As Tables VII and VIII indicate, increasing the number of windows could enhance the classification accuracy. For example in Table VII, the classification accuracy of composite of group G1 was increased from 88% to 100% by choosing nw = 35.
578
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 19, NO. 2, MARCH 2015
TABLE VII EFFECT OF THE TEMPORAL VOTING WITH DIFFERENT NUMBER OF WINDOWS IN THE CLASSIFICATION ACCURACY OF TCN CASE
TABLE VIII EFFECT OF THE TEMPORAL VOTING WITH DIFFERENT NUMBER OF WINDOWS IN THE CLASSIFICATION ACCURACY OF TAN CASE
The maximum number of windows (44) was chosen so that the used time-interval was less than 1 s. IV. DISCUSSION AND CONCLUSION In this paper, based on the general two categories of dental restorative materials, two classification cases were studied. In the first case, tooth, composite, and noncontact classes were considered (TCN case), and in the second case, tooth, amalgam, and noncontact classes were chosen (TAN case). The cutting sounds of an air-turbine handpiece in contact with a tooth as well as two cubic samples of composite and amalgam were recorded. The averaged absolute values of discrete STFT coefficients were selected as the features, and the SVM was used for classification. Different and separate groups of training and testing sets were considered for SVM as shown in Table I. During this study, the classifier performance was evaluated from different aspects including number of features (p), scaling
method, kernel type, and classification scheme. For TCN, it was indicated that 64 features with linear scaling, and the RBF kernel had a better performance (see Fig. 7). In this case, considering all the training/testing groups, the following accuracy ranges were obtained: tooth (74–98%), composite (80–91%), and noncontact (84–97%). For TAN, it was shown that 32 features with normal scaling, and the linear kernel resulted in a better accuracy (see Fig. 8). In this case, the accuracy ranges were 92–97% for tooth, 83–97% for amalgam, and 85–95% for noncontact class. According to Figs. 7 and 8, the total classification accuracies were initially enhanced by increasing the number of features (p), and then reduced for higher values of p. This behavior is referred to as “peaking phenomenon,” and has been recognized and studied for more than 60 years [38]–[40]. “Peaking phenomenon” states that for a fixed sample size (that holds true in our study), the classification accuracy increases and then decreases as the number of features grows (a peak occurs) [39]. In this study, more features could increase the separation among
ZAKERI et al.: DISCRIMINATION OF TOOTH LAYERS AND DENTAL RESTORATIVE MATERIALS USING CUTTING SOUNDS
classes; however, because of the limited size of data, the added features eventually acted similar to noises, and reduced the accuracies. In this study, the experiments were conducted offline. However, our method is applicable for real-time implementations, because of its appropriate windowing and computational complexity. The proposed method had a short WL, which made extracting the features as fast as possible. In addition, in our application, the training stage can be conducted offline. Therefore, it would not add any computational complexity during a real-time implementation. After the training, the classifier can be used to discriminate the cutting sounds (testing stage). The computational complexity of the classifier during testing stage is low, particularly for the linear kernel which is in the order of p (number of features) [27], and therefore it is suitable for real-time applications. The effect of the hierarchical classification scheme on the classification accuracy was also studied, and the results were compared with the OAO scheme. In TCN, the total accuracy for the hierarchical and OAO schemes was 88% and 89%, respectively, and for TAN, it was 91% (hierarchical), and 92% (OAO). The concept of temporal voting was also introduced, and its classification accuracy was studied using different number of windows as 2, 5, 10, 15, 20, 25, 30, 35, 40, and 44. The maximum number of windows (44) was chosen so that the time interval was less than 1 s. In TCN, the classification accuracy of all classes in all training/testing groups was improved using a temporal voting technique. For example, 44 windows resulted in the classification accuracy ranges as tooth (86–100%), composite (90–100%), and noncontact (93–98%). In TAN, except amalgam class of one group, the classification accuracy of all classes in all training/testing groups was improved using the temporal voting technique. For example, 44 windows resulted in the classification accuracy ranges as tooth (98–100%), amalgam (83–100%), and noncontact (85–97%). It should be emphasized that the time-interval equal to 1 s was selected to show that we could improve the accuracies to the above ranges; however, if the time interval is identified as critical, it can be set to any other value. The obtained classification accuracies were very promising, which provide a solid ground for further investigations. This study indicated the feasibility and capability of the proposed method in discrimination between tooth layers and restorative materials. Ultimately, the final application of this technology will be preventing inadvertent tooth layers removal during complex restorative procedures. Moreover, this innovative method can assist dentists and dental students in discerning the boundary of the tooth and restorative materials objectively. In particular, this method has great teaching values in dental educational institutes where dental students develop their tactile and sensory experience practicing on models, extracted teeth and patients. REFERENCES [1] R. G. Craig and M. L. Ward, Restorative Dental Materials, 10th ed. Louis, MO, USA: Mosby, 1997.
St.
579
[2] I. A. Mjor and A. Jokstad, “Five-year study of class II restorations in permanent teeth using amalgam, glass polyalkenoate (ionomer) cement and resin-based composite materials,” J. Dentistry, vol. 21, pp. 338–343, 1993. [3] C. J. Collins, R. W. Bryant, and K. L. Hodge, “A clinical evaluation of posterior composite resin restorations: 8-year findings,” J. Dentistry, vol. 26, pp. 311–317, 1998. [4] J. A. Soncini, N. N. Maserejian, F. Trachtenberg, M. Tavares, and C. Hayes, “The longevity of amalgam versus compomer/composite restorations in posterior primary and permanent teeth,” J. Amer. Dental Assoc., vol. 138, no. 6, pp. 763–772, 2007. [5] G. H. Johnson, D. J. Bales, G. E. Gordon, and L. V Powell, “Clinical performance of posterior composite resin restorations,” Quintessence Int., vol. 23, no. 10, pp. 705–711, Oct. 1992. [6] C. W. Drake, “A comparison of restoration longevity in maxillary and mandibular teeth,” J. Amer. Dental Assoc., vol. 116, no. 6, pp. 651–654, May 1988. [7] A. Jokstad and I. A. Mj¨or, “Replacement reasons and service time of class-II amalgam restorations in relation to cavity design,” Acta Odontol. Scand., vol. 49, no. 2, pp. 109–126, Apr. 1991. [8] M. J. Wahl, M. M. Schmitt, D. A. Overton, and M. K. Gordon, “Prevalence of cusp fractures in teeth restored with amalgam and with resin-based composite,” J. Amer. Dental Assoc., vol. 135, no. 8, pp. 1127–1132, Aug. 2004. [9] J. C. Setcos, R. Khosravi, N. H. F. Wilson, C. Shen, M. Yang, and I. A. Mj¨or, “Repair or replacement of amalgam restorations: Decisions at a USA and a UK dental school,” Operative Dentistry, vol. 29, no. 4, pp. 392–397, 2004. [10] A. O. Adegbembo and P. A. Watson, “Removal, replacement and placement of amalgam restorations by Ontario dentists in 2002,” J. Can. Dental Assoc., vol. 71, no. 8, pp. 565–565, Sep. 2005. [11] F. Sardenberg, C. C. Bonif´acio, M. M. Braga, J. C. P. Imparato, and F. M. Mendes, “Evaluation of the dental structure loss produced during maintenance and replacement of occlusal amalgam restorations,” Braz. Oral Res., vol. 22, no. 3, pp. 242–246, 2008. [12] J. D. Bader and D. A. Shugars, “What do we know about how dentists make caries-related treatment decisions?,” Community Dentistry Oral Epidemiol., vol. 25, no. 1, pp. 97–103, Feb. 1997. [13] V. V. Gordan, C. W. Garvan, J. S. Richman, J. L. Fellows, D. B. Rindal, V. Qvist, M. W. Heft, O. D. Williams, and G. H. Gilbert, “How dentists diagnose and treat defective restorations: Evidence from the dental practicebased research network,” Operative Dentistry, vol. 34, no. 6, pp. 664–673, 2009. [14] V. V. Gordan, “In vitro evaluation of margins of replaced resin-based composite restorations,” J. Esthetic Dentistry, vol. 12, no. 4, pp. 209–215, Jan. 2000. [15] V. V. Gordan, “Clinical evaluation of replacement of class V resin based composite restorations,” J. Dentistry, vol. 29, no. 7, pp. 485–488, Sep. 2001. [16] V. V. Gordan, E. Mondragon, and C. Shen, “Replacement of resin-based composite: Evaluation of cavity design, cavity depth, and shade matching,” Quintessence Int., vol. 33, no. 4, pp. 273–278, Apr. 2002. [17] T. Kocher, J. Strackeljan, and D. Behr, “Feasibility of computer-assisted recognition of different dental hard tissues,” J. Dental Res., vol. 79, no. 3, pp. 829–834, Mar. 2000. [18] H. Trabelsi and E. Kannatey-Asibu Jr., “Pattern-recognition analysis of sound radiation in metal cutting,” Int. J. Adv. Manuf. Technol., vol. 6, no. 3, pp. 220–231, 1991. [19] C. Wan, A. Mita, and T. Kume, “An automatic pipeline monitoring system using sound information,” Struct. Control Health Monit., vol. 17, no. 1, pp. 83–97, Feb. 2010. [20] S. K. Yadav, K. Tyagi, B. Shah, and P. K. Kalra, “Audio signature-based condition monitoring of internal combustion engine using FFT and correlation approach,” IEEE Trans. Instrum. Meas., vol. 60, no. 4, pp. 1217– 1226, Apr. 2011. [21] O. Amft, M. Kusserow, and G. Tr¨oster, “Bite weight prediction from acoustic recognition of chewing,” IEEE Trans. Biomed. Eng., vol. 56, no. 6, pp. 1663–1672, Jun. 2009. [22] S.-H. Shin, T. Hashimoto, and S. Hatano, “Automatic detection system for cough sounds as a symptom of abnormal health condition,” IEEE Trans. Inf. Technol. Biomed., vol. 13, no. 4, pp. 486–493, Jul. 2009. [23] A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing, 3rd ed. Englewood Cliffs, NJ, USA: Prentice-Hall, 2009. [24] V. Zakeri and S. Arzanpour, “A speed regulating scheme for air-turbine dental handpieces,” in Proc. Amer. Control Conf., 2012, pp. 2356–2361.
580
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 19, NO. 2, MARCH 2015
[25] J. O. Smith, “Spectral Audio Signal Processing,” W3K Publishing, 2011. [Online]. Available: https://ccrma.stanford.edu/∼jos/sasp/ [26] C. Hsu, C. Chang, and C. Lin, “A practical guide to support vector classification,” Dept. Comput. Sci., Nat. Taiwan Univ., 2003. [27] C. M. Bishop, Pattern Recognition and Machine Learning. New York, NY, USA: Springer, 2007. [28] J. C. Wang, J. F. Wang, K. W. He, and C. S. Hsu, “Environmental sound classification using hybrid SVM/KNN classifier and MPEG-7 audio lowlevel descriptor,” in Proc. IEEE Int. Joint Conf. Neural Netw., 2006, pp. 1731–1735. [29] J. C. Wang, H. P. Lee, J. F. Wang, and C. B. Lin, “Robust environmental sound recognition for home automation,” IEEE Trans. Autom. Sci. Eng., vol. 5, no. 1, pp. 25–31, Jan. 2008. [30] S. Choi and Z. Jiang, “Cardiac sound murmurs classification with autoregressive spectral analysis and multi-support vector machine technique,” Comput. Biol. Med., vol. 40, no. 1, pp. 8–20, 2010. [31] Z. Jiang, S. Choi, and H. Wang, “A new approach on heart murmurs classification with SVM technique,” in Proc. Int. Symp. Inf. Technol. Convergence, 2007, pp. 240–244. [32] A. Ganapathiraju, J. Hamaker, and J. Picone, “Hybrid SVM/HMM architectures for speech recognition,” in Proc. Int. Conf. Spoken Language Process., 2000, vol. 4, pp. 504–507. [33] Y.-L. Lin and G. Wei, “Speech emotion recognition based on HMM and SVM,” in Proc. Int. Conf. Mach. Learning Cybern., 2005, vol. 8, pp. 4898– 4901. [34] H. Yu and S. Kim, “SVM tutorial: Classification, regression, and ranking,” in Handbook of Natural Computing. New York, NY, USA: Springer, 2009. [35] T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical Learning: Data Mining, Inference, and Prediction, 2nd ed. New York, NY, USA: Springer, 2009. [36] C. W. Hsu and C. J. Lin, “A comparison of methods for multiclass support vector machines,” IEEE Trans. Neural Netw., vol. 13, no. 2, pp. 415–425, Jan. 2002. [37] C. C. Chang and C. J. Lin, “LIBSVM: A library for support vector machines,” Trans. Intell. Syst. Technol., vol. 2, no. 3, pp. 27:1–27:27, 2011. [38] C. Rao, “On some problems arising out of discrimination with multiple characters,” Sankhy¯a Indian J. Stat., vol. 9, no. 4, pp. 343–366, 1949. [39] A. Navot, R. Gilad-Bachrach, Y. Navot, and N. Tishby, “Is feature selection still necessary?,” in Subspace, Latent Structure and Feature Selection, vol. 3940, C. Saunders, M. Grobelnik, S. Gunn, and J. Shawe-Taylor, Eds., Berlin, Germany: Springer, 2006, pp. 127–138. [40] C. Sima and E. R. Dougherty, “The peaking phenomenon in the presence of feature-selection,” Pattern Recognit. Lett., vol. 29, pp. 1667–1674, 2008.
Vahid Zakeri (M’09) received the B.Sc. and M.Sc. degrees in electrical engineering from Shiraz University, Shiraz, Iran, in 2005, and 2008, respectively, and the Ph.D. degree in mechatronic systems engineering from Simon Fraser University, Surrey, BC, Canada, in 2013. He is currently a Postdoctoral Fellow at Simon Fraser University. The main themes of his research are biomedical engineering, signal processing, and control design. His current research is focused on developing medical assisting devices.
Siamak Arzanpour received the B.Sc. degree from the University of Tehran, Tehran, Iran, in 1998, the M.Sc. degree from the University of Toronto, Toronto, ON, Canada, in 2003, and the Ph.D. degree from the University of Waterloo, Waterloo, ON, in 2006, all in mechanical engineering. He is currently an Associate Professor at Simon Fraser University, Surrey, BC, Canada. His current research interests include a wide range of topics, including smart materials, vibration, haptic systems, pattern and material recognition using vibration signatures of biomaterials, and energy harvesting from mechanical vibrations for remote sensors.
Babak Chehroudi completed his dental doctorate training at the National University of Iran in 1983. He then started research in the area of cellular interactions with biomaterials and received the Ph.D. degree from the University of British Columbia (UBC), Vancouver, BC, Canada, in 1991. He is currently a Clinical Professor in the Department of Oral Health Sciences at the UBC Faculty of Dentistry. His research interests include interactions of biomaterials with living tissues, osseointegration of dental implants, dental implant stability measurements, virtual reality and computer-assisted teaching in dental anatomy and restorative dentistry. He also maintains a private dental practice in North Vancouver.
571
Discrimination of Tooth Layers and Dental Restorative Materials Using Cutting Sounds Vahid Zakeri, Member, IEEE, Siamak Arzanpour, and Babak Chehroudi
Abstract—Dental restoration begins with removing carries and affected tissues with air-turbine rotary cutting handpieces, and later restoring the lost tissues with appropriate restorative materials to retain the functionality. Most restoration materials eventually fail as they age and need to be replaced. One of the difficulties in replacing failing restorations is discerning the boundary of restorative materials, which causes inadvertent removal of healthy tooth layers. Developing an objective and sensor-based method is a promising approach to monitor dental restorative operations and to prevent excessive tooth losses. This paper has analyzed cutting sounds of an air-turbine handpiece to discriminate between tooth layers and two commonly used restorative materials, amalgam and composite. Support vector machines were employed for classification, and the averaged short-time Fourier transform coefficients were selected as the features. The classifier performance was evaluated from different aspects such as the number of features, feature scaling methods, classification schemes, and utilized kernels. The total classification accuracies were 89% and 92% for cases included composite and amalgam materials, respectively. The obtained results indicated the feasibility and effectiveness of the proposed method.
Fig. 1. Two commonly used dental restorations; (a) amalgam; (b) composite; (c) fractured (broken) composite which should be replaced.
Index Terms—Audio monitoring, audio signal processing, dental restoration, support vector machine (SVM).
I. INTRODUCTION OOTH is an inhomogeneous structure which is composed of different layers including enamel, dentin, pulp, and cementum. Tooth decay is one of the most prevalent disorders, and dentists need to remove infected parts to restore tooth functionality. Dental restoration is a process that begins with removing carries and affected tissues to retain the functionality of tooth layers. Air-turbine dental handpieces are high-speed rotary cutting tools that are widely used by dentists during this operation. The next stage in the process is filling the cavity with appropriate restorative materials. There are different materials such as amalgam, composite, glass ionomer, ceramics, and gold that are
T
Manuscript received July 22, 2013; revised December 16, 2013 and March 27, 2014; accepted April 4, 2014. Date of publication April 15, 2014; date of current version March 2, 2015. This work was supported by funds provided through the discovery research grant of Natural Sciences and Engineering Research Council of Canada, and Sable Industries, Inc. V. Zakeri and S. Arzanpour are with the School of Mechatronic Systems Engineering, Simon Fraser University, Surrey, BC V3T 0A3, Canada (e-mail: [email protected]; [email protected]). B. Chehroudi is with the Faculty of Dentistry, University of British Columbia, Vancouver, BC V6T 1Z4, Canada (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/JBHI.2014.2317503
commercially available for dental restorations [1]. Among these materials, amalgam and composite are widely used by dentists (see Fig. 1). The main advantage of composite over amalgam is its aesthetical feature, because it can be produced in a variety of tooth-matching shades, which makes its appearance very similar to a real tooth. Although current dental restoration materials are very durable, external mechanical forces of clenching and grinding, as well as chronic exposure to acids and solvents in the mouth may result in fatigue, cracks and ultimately failure [see Fig. 1(c)]. Moreover, the performance of dental restorations is also subject to several factors such as restorative materials [2]–[4], the type and position of the tooth [5], [6], the restoration’s shape, size, and number of involved surfaces [7], [8], and the patient’s age [2], [8]. If the old restoration collapses, there is a high potential for developing new decay that requires replacing of the old restoration and removal of all carries. Replacing old failing restorations is one of the most frequent procedures performed in dental practices [9]–[11]. This rate has not been declined despite of all the advancements in dental restoration materials. To conduct dental restorative operations, dentists receive extensive trainings to become experts in interpreting their tactile, auditory, and visual senses as their diagnosis tools [12]. They transfer such sensory information to the actual practice through their perception. This perceptual procedure is highly subjective, and depends on the individual abilities and experience [13]. The main problem with these approaches is that human senses have
2168-2194 © 2014 Crown Copyright
572
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 19, NO. 2, MARCH 2015
limited functionality, and often are insufficient for dentists to rely on. For instance, discerning the boundary of restoration is challenging, in particular with tooth-colored composite restorations. In fact, during the removal of old composite restorations, healthy tooth layers in adjacent areas are often inadvertently removed [13]–[16]. Although the visibility issue is less challenging for amalgam, replacing it still results in tooth losses [11]. Loosing healthy parts of a tooth is a major problem, because it reduces the longevity of the restoration [15]. In addition, tooth is one of the few human body parts that has very limited healing ability, and almost all losses will be permanent. Therefore, developing an objective and sensor-based method for dental restorative operations is a promising approach to overcome the limited functionality of human senses. The main purpose of this approach is to discriminate restorative materials and tooth layers during the cutting and removal procedures. To achieve this goal, this paper studies the feasibility of analyzing cutting sounds of restoration processes as a tool to discriminate tooth layers and restorative materials. The difference among the mechanical properties of tooth layers and restorative materials (e.g., hardness, density, elasticity, etc.) [1], [17] implies that their cutting sounds should carry unique features that might be useful to accurately discriminate different materials. Moreover, the cutting sound is a valuable source of information that is readily and easily available during the cutting procedure. Thus far, no prior works have been found on discrimination of tooth and restorative materials based on the cutting sounds. Perhaps the closest study to this work was conducted by Kocher et al. [17]. They used an ultrasonic dental scaler, and measured its oscillations in contact with a tooth. They successfully validated their approach offline, and recognized enamel, cementum, and calculus tissues. However, no information was found in their report on recognizing the noncontact class (i.e., when the tip of the scaler did not have contact with the tooth). As it will be discussed in the next section, it is necessary in our study to define and recognize such a class. In a broad perspective, our proposed method falls into the field of audio-based monitoring. Audio signals have been studied and used in various applications for monitoring purposes in both real-time and offline settings. Trabelsi et al. [18] proposed a tool-wear monitoring procedure based on the sounds generated in metal cuttings. They followed offline approach to discriminate between sharp, worn, and broken tools. Wan et al. [19] developed an automatic pipeline monitoring system that used sound information of road cutters to avoid accidental breakage. The feasibility of their approach was investigated offline, and it was concluded that further tests would be needed for the on-site environments. In another study, Yadav et al. [20] introduced a real-time condition monitoring system based on audio signals for internal combustion engines. They could effectively identify faults within the engine, and differentiate them from healthy conditions. Audio signals have also been employed in dietary and health monitoring systems. Amft et al. [21] presented an automatic
Fig. 2. Proposed methodology for TCN and TAN classification cases; TCN: tooth/composite/noncontact; TAN: tooth/amalgam/noncontact; STFT: shorttime Fourier transform; SVM: support vector machine; RBF: radial basis function.
procedure to predict food weight based on acoustic recordings of chewing. They conducted the tests offline, and concluded that their approach was feasible for solid foods. In another work, Shin et al. [22] developed an automatic system for real-time monitoring of health conditions using cough sounds. In this paper, we present a novel monitoring system for dental restorative operations based on the cutting sounds that can discriminate tooth layers and restorative materials. The composition of this paper is as follows: a systematic theoretical– experimental methodology is proposed in Section II, Section III presents and elaborates the results, and finally, Section IV provides discussions and conclusions. II. MATERIAL AND METHOD Considering two widely used restorative materials, i.e., amalgam and composite, and the fact that a tooth is not filled with both materials simultaneously, we are facing two cases of classification: 1) Case 1: Classification between tooth, composite, and noncontact classes (TCN), 2) Case 2: Classification between tooth, amalgam, and noncontact classes (TAN). It should be noted that noncontact was considered as a “class,” because the classifier should be able to identify the free-running handpiece, and discriminate the sound from those of the contact classes (tooth or composite/amalgam). To investigate TCN and TAN classification cases, a methodology is proposed which is depicted in Fig. 2. In this methodology, seven different stages were considered as follows. A. Data Collection For data collection, an experienced dentist while maintaining similar conditions to actual cutting procedures conducted several tests in a dental clinic. In vitro tests were performed on cubic samples (1 cm ×1 cm ×1 cm) of amalgam and composite, as well as one intact extracted human third molar. The samples were fixed in a metal chuck (clamp) during the cutting procedure. A “W & H Toplight 898le” air-turbine handpiece and a “330 Diamond” bur were used for cutting, which are among common choices for dentists in restorative procedures.
ZAKERI et al.: DISCRIMINATION OF TOOTH LAYERS AND DENTAL RESTORATIVE MATERIALS USING CUTTING SOUNDS
Fig. 3.
(a) Composite, amalgam and tooth samples; (b) metal chuck (clamp) for holding the cutting samples; (c) microphone and air-turbine handpiece.
Fig. 4.
Sample STFT spectrum (noncontact): (a) before filtering, and (b) after filtering.
A high-frequency microphone (GRAS 40be) was attached to the handpiece to record the cutting sounds (see Fig. 3).This arrangement ensured that the relative position of the handpiece and the microphone was fixed for all tests. The sampling frequency was 48 kHz which was high enough to capture the maximum speed of the handpiece (∼5000 r/s) based on Nyquist–Shannon sampling theorem [23]. A high-speed data acquisition card (LDS Dactron Photon II), and a signal processing software for frequency analysis (RT Pro Photon) were employed to sample and collect generated data on a personal computer. In obtaining the tooth data, the cutting sounds of enamel and dentin layers were recorded. In addition, similar to the actual restorative process, the cutting sounds were recorded with/without water spray (“wet” and “dry”). All cuttings were conducted three times in a parallel plane to the samples’ surface (long axis of the cutting bur was parallel with the long axis of the tooth). The cutting procedure was maintained constant for all samples, and it comprised of three steps. First, the handpiece ran freely for 3 s (noncontact 1) and then the cutting was undertaken for 3 s (contact). In the last step (noncontact 2), the handpiece ran freely again for 2 s. In this cutting procedure, the transition of a noncontact sound to a contact sound and vice versa were considered. B. Preprocessing and Labeling At this stage, the recorded data were filtered using a highpass filter (Butterworth, order 3) with the cut-off frequency at
573
4 kHz to remove the effect of noises and disturbances (e.g., the compressor sound) that happened at lower frequencies. Fig. 4 provides a comparison of a noncontact sample data before [see Fig. 4(a)], and after the filtering [see Fig. 4(b)]. As it can be seen, the filter could efficiently remove the disturbance sound [the first peak in Fig. 4(a)]. The higher peak (around 5000 Hz) indicates the sound pitch of the handpiece when it was running free (noncontact). In each test, the cutting segment was identified manually based on an increase to the signal power and a decrease in the handpiece’s speed. The speed was measured by the method described in [24]. The cutting data then labeled appropriately as tooth, composite, or amalgam. The free-running data of all tests were labeled as noncontact. It should be mentioned that in the labeling process, both dry and wet cutting signals were considered together for each material to include pertinent information. Fig. 5 shows a sample of a recorded sound signal. The cutting period is highlighted by the rectangle, and the rest of data are noncontact. C. Windowing, Feature Extraction, and Feature Scaling After preprocessing and labeling, the next stages of the proposed methodology are windowing, feature extraction, and feature scaling (see Fig. 2).The absolute value of discrete short-time Fourier transform (STFT) coefficients [23] were selected as the features. Hamming windowing with 50% of overlapping was used, because it is a common choice in audio signal processing [25]. The window length (WL) should be chosen small,
574
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 19, NO. 2, MARCH 2015
D. Classification
Fig. 5. Sample of a recorded sound signal; the cutting period is highlighted by the rectangle, and the rest of data are noncontact.
To deal with the classification problem, the machine learning approach was selected. This approach focuses on developing and studying the models that can learn from data [27]. Support vector machine (SVM) is one of the most powerful machine learning algorithms that was chosen in this study. SVM is a popular method for classification that has been employed in various applications such as environmental sound classification [28], [29], cardiac sound classification [30], [31], and speech recognition [32], [33]. SVM is originally a classifier that discriminates data points of two classes by finding a hyperplane that maximizes the separation between the two classes [34]. Each data point belongs to only one class, and is represented by an n-dimensional vector (xi ∈ Rn , i = 1, . . . , m). Each xi is assigned a class-label yi . For a two-class problem, the labels are considered either 1 or −1 (yi ∈ {1, − 1}). The hyperplane takes the format of f (x) = wT x + b
Fig. 6.
Sample STFT spectrum and the averaged coefficients for p = 32.
because we are interested in discrimination between the classes in a small period to be able to prevent the removal of healthy tooth layers. For this study, 2100 data samples were selected in each window, and since the sampling frequency was 48 kHz, the WL was 2100/48 000 ∼ 44 ms. For each window, 2048 Fourier transform coefficients were obtained using the fast Fourier transform algorithm. Since each feature vector had a high dimension (2048 elements), the computation cost could have been increased. To reduce the dimension of each feature vector, a mathematical averaging technique was used. In this technique, 2048 STFT coefficients were divided coefficients in each group). The to p equal-sized groups ( 2048 p data of each group were averaged mathematically to obtain p features. Fig. 6 indicates a sample STFT spectrum and the averaged coefficients for p = 32. The effect of p on the classification accuracy will be explored in the “result” section. Feature scaling is usually utilized to prevent attributes in greater numeric ranges dominate those in smaller numeric ranges [26]. In addition, it can decrease numerical difficulties during the calculation. For this study, two feature scaling methods were investigated: linear scaling and normal scaling. In the linear scaling, all the vector elements were projected linearly to the interval [0, 1], and in the normal scaling, the vector elements were normalized to have a zero mean and unit variance. The effect of these scaling methods on the classification accuracy will be discussed in the “result” section.
(1)
where w and b are the weight vector and the bias (wT denotes the transpose of w). Finding the values of w and b is referred to as the training process, in which SVM uses a selective set of data points that are denoted as training data. After training, another set of data points can be used to evaluate the performance of SVM (testing data). In this study, two different sets of data were used for training and testing. In order to correctly classify the training data, the hyperplane f should be positive for positive data (yi = 1), and negative for negative data (yi = −1). It can be shown that SVM would be formulated as a constrained optimization problem [27] 1 w 2 +C εi 2 i subject to yi wT xi + b ≥ 1 − εi
minw ,ε
(2) εi ≥ 0
(3)
where ε is a slack variable that measures the degree of misclassification, and C is a “hyperparameter” that compromises between classes’ separation and the amount of misclassification. The above-constrained optimization problem can be solved using Lagrange multipliers [27]. One approach for the data points that are not separable in their original space x is to transform them to a new space φ (x) that usually has a higher dimension. In the new space φ (x), the transformed data points will become separable. To reduce the computation cost of a higher dimension space, a technique called kernel method can be employed. The idea of this method is that for some new spaces φ (x), the dot product between the transformed data points (φ(xi ).φ (xj )) is computed efficiently. It can be shown that the solution of (2) and (3) can be rewritten so that it has only dot products of data points (xi .xj ) [27]. Therefore, when data points are transformed to a new space φ (x), the new solution only depends on the kernel function Ker (xi , xj ) = φ(xi ).φ (xj ), and the kernel method is applicable.
ZAKERI et al.: DISCRIMINATION OF TOOTH LAYERS AND DENTAL RESTORATIVE MATERIALS USING CUTTING SOUNDS
575
TABLE I DIFFERENT GROUPS FOR TRAINING AND TESTING
Two commonly used kernels for SVM are radial basis function (RBF) kernel and linear kernel which are described by [27] (4) K (xi , xj ) = exp −γ xi − xj 2 ; γ > 0 and K (xi , xj ) = xi .xj .
(5)
According to (5), in the linear kernel, the original space x and the new space φ (x) are identical. In (4), γ is a hyper-parameter [similar to C in (2)]. To find the optimal values of these hyperparameters, usually a “grid search” approach is employed, in which the values of the hyperparameters are changed, and the optimal values are found based on the cross validation [26]. In a ν-fold cross validation, the training data points are divided to ν equal-sized subsets. Then, ν − 1 subsets are used for training, and the one remained subset is utilized for testing. The entire procedure is performed ν times, and the total classification error is obtained [35]. Those values of C and γ will be selected that result in the minimum error. In this study, the hyperparameters were varied from 2−15 , 2−13 , . . . to 21 , 23 ; and a fivefold cross validation was employed. SVM is basically designed for binary classifications. Therefore, we need to come up with methods to use it for our application, which is a three-class problem. Hsu and Lin compared different methods for multiclass SVM, and concluded that a “one-against-one” (OAO) scheme was more suitable considering practical aspects [36]. In this scheme, β (β − 1) /2 binary classifiers are trained (β is the number of classes), and the class that receives the maximum “votes” is assigned to a test data. III. RESULT A. Selecting the Number of Features, Scaling Method, and SVM Kernel As it was explained in the previous section, SVM first should be trained. Two of the three experiments were used for training, and the third one was employed for testing. Table I indicates different groups of training and testing sets. To apply SVM, the software LIBSVM was used [37]. Considering all classes/groups, the total accuracy of the classifier was obtained for each case (TCN or TAN). As it was mentioned in the “classification” section, the values of hyperparameters [C in (2) and γ in (4)] were found by a fivefold cross validation. The classifier performance was evaluated from different aspects. The number of features (p) was varied as 2, 4, 8, 16, 32, 64, 128, 256, 512, and for each number, the effect of linear/normal scaling methods and linear/RBF kernels on the total accuracy was studied. Figs. 7 and 8 show the results for TCN and TAN cases, respectively.
Fig. 7. Effect of different number of features, scaling methods, and SVM kernels on the total accuracy of TCN case (the number-of-features axis is in logarithmic scale).
Fig. 8. Effect of different number of features, scaling methods, and SVM kernels on the total accuracy of TAN case (the number-of-features axis is in logarithmic scale).
As Fig. 7 indicates for TCN case, increasing the number of features could enhance the classification accuracy. In addition, no particular trends could be observed for different scaling methods and kernels. In fact, depending on the value of p, one selection of scaling and kernel could perform better than others. For example, for p = 16, linear scaling/linear kernel had the maximum accuracy, while for p = 512 normal scaling/RBF kernel was maximum. In general, the maximum accuracy was 89% which was obtained for p = 64 with linear scaling/RBF kernel, and for p = 256, 512, both with normal scaling/RBF kernel. For TAN case as Fig. 8 indicates, by increasing the number of features, the classification accuracy first reached to a maximum, stayed at that maximum for some values of p, and then eventually decreased. Generally, the accuracy order from higher to lower was obtained for normal scaling/linear kernel, linear scaling/linear kernel, linear scaling/RBF kernel, and normal scaling/RBF kernel. The maximum accuracy was 92% which was obtained for p = 32, 64, 128, all with normal scaling/linear kernel. Considering these results and the fact that less number of features causes less computation costs, 64 features with linear scaling/RBF kernel were selected for TCN, whereas 32 features with normal scaling/linear kernel were selected for TAN. Table II displays the hyperparameters’ values of these selections.
576
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 19, NO. 2, MARCH 2015
TABLE II HYPERPARAMETERS’ VALUES FROM FIVE FOLD CROSS VALIDATION
TABLE III CLASSIFICATION ACCURACY OF TCN CASE
TABLE IV CLASSIFICATION ACCURACY OF TAN CASE
All the values of Table II resulted in 98–100% crossvalidation accuracy, and therefore it was not needed to look for other values of C and γ. In this table, no values were obtained for γ in TAN case, because a linear kernel was employed. The hyperparameters of Table II were used to train the SVM classifiers for TCN and TAN cases. Table III (TCN) and Table IV (TAN) indicate the accuracy of these classifiers on the testing data of each group. According to Table III, the worst accuracy for TCN case was 74% which was obtained in classifying of 608 test samples of tooth data in group G2. The best accuracy for this case was for tooth class of group G3 with 98% correct classification out of 567 test samples. For TAN case as Table IV indicates, the worst accuracy was for amalgam of group G1 with 83% correct classification out of 368 test samples. The best accuracy for this case was for both tooth and amalgam classes of group G3 with 97% correct classification out of 567 and 293 test samples, respectively. B. Classification Scheme As it was mentioned in the “classification” section, the OAO scheme was used for SVM. An alternative approach is
Fig. 9.
Hierarchical classification scheme.
considering a hierarchical classification scheme. In this scheme, as indicated by Fig. 9, two levels are considered. In the first level, the classifier decides if the data belong to the contact or noncontact classes. The second level is used for contact data, and based on the classification case, it determines between tooth/composite [case 1 in Fig. 9], or tooth/amalgam [case 2 in Fig. 9]. In order to employ the hierarchical scheme for TCN case, two binary class models for contact/noncontact (Con/N) and tooth/composite (T/C) were obtained. To train these models, similar to OAO scheme, 64 features with linear scaling/RBF kernel were used. In TAN case, two binary class models were obtained for contact/noncontact (Con/N) and tooth/amalgam (T/A). To train these models, 32 features with normal scaling/linear kernel were utilized. In the hierarchical scheme, the values of hyperparameters for both cases were also found by a fivefold cross validation. The developed binary models were employed in the hierarchical scheme (see Fig. 9). Table V (TCN) and Table VI (TAN) show the classification accuracy of this scheme and compare the results with the OAO scheme. As Table V shows, the hierarchical scheme improved the classification accuracy of the tooth class for groups G1 and G2. The accuracy in classification of composite of G2 as well as noncontact classes of G2 and G3 were also increased for the hierarchical scheme. According to Table VI, tooth class of group G2, amalgam class of group G1, and noncontact class of groups G1 and G3, had a better accuracy in the hierarchical scheme. For other classes/groups, the OAO scheme had better performance. The maximum difference between the accuracies of both schemes was for tooth of G1, in which the OAO classification was 17% (96–79%) more accurate than the hierarchical scheme. Considering all groups/classes, the total accuracy in TCN was 89% and 88% for OAO and hierarchical schemes, respectively. In TAN,
ZAKERI et al.: DISCRIMINATION OF TOOTH LAYERS AND DENTAL RESTORATIVE MATERIALS USING CUTTING SOUNDS
577
TABLE V COMPARISON BETWEEN OAO AND HIERARCHICAL CLASSIFICATION SCHEMES FOR TCN CASE
TABLE VI COMPARISON BETWEEN OAO AND HIERARCHICAL CLASSIFICATION SCHEMES FOR TAN CASE
the total accuracy for OAO classification was 92% while for the hierarchical scheme was 91%. C. Temporal Voting Temporal voting is a promising approach to enhance the classification accuracy. As it was mentioned previously in Section II, the recorded sounds were windowed with 50% overlapping. Then, the features were extracted from each window and considered as one sample data. In this approach, the used time-interval was equal to the WL as Fig. 10(a) shows. For this study, the WL was WL = 482100 000 . In the temporal voting approach, the time interval is chosen so that it covers more than one window. As Fig. 10(b) indicates, the time interval is corresponding to five windows (considering 50% of overlapping). In this approach, the class of each sample in the selected time interval is first found separately, e.g., the class of data associated with W1, W2, . . . ,W5. Then, the class that receives the maximum votes will be assigned to the sample associated with W1. If the selected number of windows in the temporal voting approach is nw (with 50% overlapping) and the length of each window is WL, the used time-interval is (nw + 1) WL . 2
(6)
Fig. 10.
Schematic diagram of the temporal voting approach.
To employ temporal voting approach, TCN and TAN classifiers of the OAO scheme were used. Tables VII and VIII compare the effect of temporal voting with different number of windows (nw ) on the classification accuracies of both cases. As Tables VII and VIII indicate, increasing the number of windows could enhance the classification accuracy. For example in Table VII, the classification accuracy of composite of group G1 was increased from 88% to 100% by choosing nw = 35.
578
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 19, NO. 2, MARCH 2015
TABLE VII EFFECT OF THE TEMPORAL VOTING WITH DIFFERENT NUMBER OF WINDOWS IN THE CLASSIFICATION ACCURACY OF TCN CASE
TABLE VIII EFFECT OF THE TEMPORAL VOTING WITH DIFFERENT NUMBER OF WINDOWS IN THE CLASSIFICATION ACCURACY OF TAN CASE
The maximum number of windows (44) was chosen so that the used time-interval was less than 1 s. IV. DISCUSSION AND CONCLUSION In this paper, based on the general two categories of dental restorative materials, two classification cases were studied. In the first case, tooth, composite, and noncontact classes were considered (TCN case), and in the second case, tooth, amalgam, and noncontact classes were chosen (TAN case). The cutting sounds of an air-turbine handpiece in contact with a tooth as well as two cubic samples of composite and amalgam were recorded. The averaged absolute values of discrete STFT coefficients were selected as the features, and the SVM was used for classification. Different and separate groups of training and testing sets were considered for SVM as shown in Table I. During this study, the classifier performance was evaluated from different aspects including number of features (p), scaling
method, kernel type, and classification scheme. For TCN, it was indicated that 64 features with linear scaling, and the RBF kernel had a better performance (see Fig. 7). In this case, considering all the training/testing groups, the following accuracy ranges were obtained: tooth (74–98%), composite (80–91%), and noncontact (84–97%). For TAN, it was shown that 32 features with normal scaling, and the linear kernel resulted in a better accuracy (see Fig. 8). In this case, the accuracy ranges were 92–97% for tooth, 83–97% for amalgam, and 85–95% for noncontact class. According to Figs. 7 and 8, the total classification accuracies were initially enhanced by increasing the number of features (p), and then reduced for higher values of p. This behavior is referred to as “peaking phenomenon,” and has been recognized and studied for more than 60 years [38]–[40]. “Peaking phenomenon” states that for a fixed sample size (that holds true in our study), the classification accuracy increases and then decreases as the number of features grows (a peak occurs) [39]. In this study, more features could increase the separation among
ZAKERI et al.: DISCRIMINATION OF TOOTH LAYERS AND DENTAL RESTORATIVE MATERIALS USING CUTTING SOUNDS
classes; however, because of the limited size of data, the added features eventually acted similar to noises, and reduced the accuracies. In this study, the experiments were conducted offline. However, our method is applicable for real-time implementations, because of its appropriate windowing and computational complexity. The proposed method had a short WL, which made extracting the features as fast as possible. In addition, in our application, the training stage can be conducted offline. Therefore, it would not add any computational complexity during a real-time implementation. After the training, the classifier can be used to discriminate the cutting sounds (testing stage). The computational complexity of the classifier during testing stage is low, particularly for the linear kernel which is in the order of p (number of features) [27], and therefore it is suitable for real-time applications. The effect of the hierarchical classification scheme on the classification accuracy was also studied, and the results were compared with the OAO scheme. In TCN, the total accuracy for the hierarchical and OAO schemes was 88% and 89%, respectively, and for TAN, it was 91% (hierarchical), and 92% (OAO). The concept of temporal voting was also introduced, and its classification accuracy was studied using different number of windows as 2, 5, 10, 15, 20, 25, 30, 35, 40, and 44. The maximum number of windows (44) was chosen so that the time interval was less than 1 s. In TCN, the classification accuracy of all classes in all training/testing groups was improved using a temporal voting technique. For example, 44 windows resulted in the classification accuracy ranges as tooth (86–100%), composite (90–100%), and noncontact (93–98%). In TAN, except amalgam class of one group, the classification accuracy of all classes in all training/testing groups was improved using the temporal voting technique. For example, 44 windows resulted in the classification accuracy ranges as tooth (98–100%), amalgam (83–100%), and noncontact (85–97%). It should be emphasized that the time-interval equal to 1 s was selected to show that we could improve the accuracies to the above ranges; however, if the time interval is identified as critical, it can be set to any other value. The obtained classification accuracies were very promising, which provide a solid ground for further investigations. This study indicated the feasibility and capability of the proposed method in discrimination between tooth layers and restorative materials. Ultimately, the final application of this technology will be preventing inadvertent tooth layers removal during complex restorative procedures. Moreover, this innovative method can assist dentists and dental students in discerning the boundary of the tooth and restorative materials objectively. In particular, this method has great teaching values in dental educational institutes where dental students develop their tactile and sensory experience practicing on models, extracted teeth and patients. REFERENCES [1] R. G. Craig and M. L. Ward, Restorative Dental Materials, 10th ed. Louis, MO, USA: Mosby, 1997.
St.
579
[2] I. A. Mjor and A. Jokstad, “Five-year study of class II restorations in permanent teeth using amalgam, glass polyalkenoate (ionomer) cement and resin-based composite materials,” J. Dentistry, vol. 21, pp. 338–343, 1993. [3] C. J. Collins, R. W. Bryant, and K. L. Hodge, “A clinical evaluation of posterior composite resin restorations: 8-year findings,” J. Dentistry, vol. 26, pp. 311–317, 1998. [4] J. A. Soncini, N. N. Maserejian, F. Trachtenberg, M. Tavares, and C. Hayes, “The longevity of amalgam versus compomer/composite restorations in posterior primary and permanent teeth,” J. Amer. Dental Assoc., vol. 138, no. 6, pp. 763–772, 2007. [5] G. H. Johnson, D. J. Bales, G. E. Gordon, and L. V Powell, “Clinical performance of posterior composite resin restorations,” Quintessence Int., vol. 23, no. 10, pp. 705–711, Oct. 1992. [6] C. W. Drake, “A comparison of restoration longevity in maxillary and mandibular teeth,” J. Amer. Dental Assoc., vol. 116, no. 6, pp. 651–654, May 1988. [7] A. Jokstad and I. A. Mj¨or, “Replacement reasons and service time of class-II amalgam restorations in relation to cavity design,” Acta Odontol. Scand., vol. 49, no. 2, pp. 109–126, Apr. 1991. [8] M. J. Wahl, M. M. Schmitt, D. A. Overton, and M. K. Gordon, “Prevalence of cusp fractures in teeth restored with amalgam and with resin-based composite,” J. Amer. Dental Assoc., vol. 135, no. 8, pp. 1127–1132, Aug. 2004. [9] J. C. Setcos, R. Khosravi, N. H. F. Wilson, C. Shen, M. Yang, and I. A. Mj¨or, “Repair or replacement of amalgam restorations: Decisions at a USA and a UK dental school,” Operative Dentistry, vol. 29, no. 4, pp. 392–397, 2004. [10] A. O. Adegbembo and P. A. Watson, “Removal, replacement and placement of amalgam restorations by Ontario dentists in 2002,” J. Can. Dental Assoc., vol. 71, no. 8, pp. 565–565, Sep. 2005. [11] F. Sardenberg, C. C. Bonif´acio, M. M. Braga, J. C. P. Imparato, and F. M. Mendes, “Evaluation of the dental structure loss produced during maintenance and replacement of occlusal amalgam restorations,” Braz. Oral Res., vol. 22, no. 3, pp. 242–246, 2008. [12] J. D. Bader and D. A. Shugars, “What do we know about how dentists make caries-related treatment decisions?,” Community Dentistry Oral Epidemiol., vol. 25, no. 1, pp. 97–103, Feb. 1997. [13] V. V. Gordan, C. W. Garvan, J. S. Richman, J. L. Fellows, D. B. Rindal, V. Qvist, M. W. Heft, O. D. Williams, and G. H. Gilbert, “How dentists diagnose and treat defective restorations: Evidence from the dental practicebased research network,” Operative Dentistry, vol. 34, no. 6, pp. 664–673, 2009. [14] V. V. Gordan, “In vitro evaluation of margins of replaced resin-based composite restorations,” J. Esthetic Dentistry, vol. 12, no. 4, pp. 209–215, Jan. 2000. [15] V. V. Gordan, “Clinical evaluation of replacement of class V resin based composite restorations,” J. Dentistry, vol. 29, no. 7, pp. 485–488, Sep. 2001. [16] V. V. Gordan, E. Mondragon, and C. Shen, “Replacement of resin-based composite: Evaluation of cavity design, cavity depth, and shade matching,” Quintessence Int., vol. 33, no. 4, pp. 273–278, Apr. 2002. [17] T. Kocher, J. Strackeljan, and D. Behr, “Feasibility of computer-assisted recognition of different dental hard tissues,” J. Dental Res., vol. 79, no. 3, pp. 829–834, Mar. 2000. [18] H. Trabelsi and E. Kannatey-Asibu Jr., “Pattern-recognition analysis of sound radiation in metal cutting,” Int. J. Adv. Manuf. Technol., vol. 6, no. 3, pp. 220–231, 1991. [19] C. Wan, A. Mita, and T. Kume, “An automatic pipeline monitoring system using sound information,” Struct. Control Health Monit., vol. 17, no. 1, pp. 83–97, Feb. 2010. [20] S. K. Yadav, K. Tyagi, B. Shah, and P. K. Kalra, “Audio signature-based condition monitoring of internal combustion engine using FFT and correlation approach,” IEEE Trans. Instrum. Meas., vol. 60, no. 4, pp. 1217– 1226, Apr. 2011. [21] O. Amft, M. Kusserow, and G. Tr¨oster, “Bite weight prediction from acoustic recognition of chewing,” IEEE Trans. Biomed. Eng., vol. 56, no. 6, pp. 1663–1672, Jun. 2009. [22] S.-H. Shin, T. Hashimoto, and S. Hatano, “Automatic detection system for cough sounds as a symptom of abnormal health condition,” IEEE Trans. Inf. Technol. Biomed., vol. 13, no. 4, pp. 486–493, Jul. 2009. [23] A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing, 3rd ed. Englewood Cliffs, NJ, USA: Prentice-Hall, 2009. [24] V. Zakeri and S. Arzanpour, “A speed regulating scheme for air-turbine dental handpieces,” in Proc. Amer. Control Conf., 2012, pp. 2356–2361.
580
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 19, NO. 2, MARCH 2015
[25] J. O. Smith, “Spectral Audio Signal Processing,” W3K Publishing, 2011. [Online]. Available: https://ccrma.stanford.edu/∼jos/sasp/ [26] C. Hsu, C. Chang, and C. Lin, “A practical guide to support vector classification,” Dept. Comput. Sci., Nat. Taiwan Univ., 2003. [27] C. M. Bishop, Pattern Recognition and Machine Learning. New York, NY, USA: Springer, 2007. [28] J. C. Wang, J. F. Wang, K. W. He, and C. S. Hsu, “Environmental sound classification using hybrid SVM/KNN classifier and MPEG-7 audio lowlevel descriptor,” in Proc. IEEE Int. Joint Conf. Neural Netw., 2006, pp. 1731–1735. [29] J. C. Wang, H. P. Lee, J. F. Wang, and C. B. Lin, “Robust environmental sound recognition for home automation,” IEEE Trans. Autom. Sci. Eng., vol. 5, no. 1, pp. 25–31, Jan. 2008. [30] S. Choi and Z. Jiang, “Cardiac sound murmurs classification with autoregressive spectral analysis and multi-support vector machine technique,” Comput. Biol. Med., vol. 40, no. 1, pp. 8–20, 2010. [31] Z. Jiang, S. Choi, and H. Wang, “A new approach on heart murmurs classification with SVM technique,” in Proc. Int. Symp. Inf. Technol. Convergence, 2007, pp. 240–244. [32] A. Ganapathiraju, J. Hamaker, and J. Picone, “Hybrid SVM/HMM architectures for speech recognition,” in Proc. Int. Conf. Spoken Language Process., 2000, vol. 4, pp. 504–507. [33] Y.-L. Lin and G. Wei, “Speech emotion recognition based on HMM and SVM,” in Proc. Int. Conf. Mach. Learning Cybern., 2005, vol. 8, pp. 4898– 4901. [34] H. Yu and S. Kim, “SVM tutorial: Classification, regression, and ranking,” in Handbook of Natural Computing. New York, NY, USA: Springer, 2009. [35] T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical Learning: Data Mining, Inference, and Prediction, 2nd ed. New York, NY, USA: Springer, 2009. [36] C. W. Hsu and C. J. Lin, “A comparison of methods for multiclass support vector machines,” IEEE Trans. Neural Netw., vol. 13, no. 2, pp. 415–425, Jan. 2002. [37] C. C. Chang and C. J. Lin, “LIBSVM: A library for support vector machines,” Trans. Intell. Syst. Technol., vol. 2, no. 3, pp. 27:1–27:27, 2011. [38] C. Rao, “On some problems arising out of discrimination with multiple characters,” Sankhy¯a Indian J. Stat., vol. 9, no. 4, pp. 343–366, 1949. [39] A. Navot, R. Gilad-Bachrach, Y. Navot, and N. Tishby, “Is feature selection still necessary?,” in Subspace, Latent Structure and Feature Selection, vol. 3940, C. Saunders, M. Grobelnik, S. Gunn, and J. Shawe-Taylor, Eds., Berlin, Germany: Springer, 2006, pp. 127–138. [40] C. Sima and E. R. Dougherty, “The peaking phenomenon in the presence of feature-selection,” Pattern Recognit. Lett., vol. 29, pp. 1667–1674, 2008.
Vahid Zakeri (M’09) received the B.Sc. and M.Sc. degrees in electrical engineering from Shiraz University, Shiraz, Iran, in 2005, and 2008, respectively, and the Ph.D. degree in mechatronic systems engineering from Simon Fraser University, Surrey, BC, Canada, in 2013. He is currently a Postdoctoral Fellow at Simon Fraser University. The main themes of his research are biomedical engineering, signal processing, and control design. His current research is focused on developing medical assisting devices.
Siamak Arzanpour received the B.Sc. degree from the University of Tehran, Tehran, Iran, in 1998, the M.Sc. degree from the University of Toronto, Toronto, ON, Canada, in 2003, and the Ph.D. degree from the University of Waterloo, Waterloo, ON, in 2006, all in mechanical engineering. He is currently an Associate Professor at Simon Fraser University, Surrey, BC, Canada. His current research interests include a wide range of topics, including smart materials, vibration, haptic systems, pattern and material recognition using vibration signatures of biomaterials, and energy harvesting from mechanical vibrations for remote sensors.
Babak Chehroudi completed his dental doctorate training at the National University of Iran in 1983. He then started research in the area of cellular interactions with biomaterials and received the Ph.D. degree from the University of British Columbia (UBC), Vancouver, BC, Canada, in 1991. He is currently a Clinical Professor in the Department of Oral Health Sciences at the UBC Faculty of Dentistry. His research interests include interactions of biomaterials with living tissues, osseointegration of dental implants, dental implant stability measurements, virtual reality and computer-assisted teaching in dental anatomy and restorative dentistry. He also maintains a private dental practice in North Vancouver.
