J Oral Maxillofac Surg 70:51-59, 2012
Use of Artificial Neural Network in Differentiation of Subgroups of Temporomandibular Internal Derangements: A Preliminary Study Burcu Bas, DDS, PhD,* Okan Ozgonenel, PhD,† Bora Ozden, DDS, PhD,‡ Burak Bekcioglu, DDS, PhD,§ Emel Bulut, DDS, PhD,储 and Murat Kurt, DDS, PhD¶ Purpose: Artificial neural networks (ANNs) have been developed in the past few decades for many
different applications in medical science and in biomedical research. The use of neural networks in oral and maxillofacial surgery is limited. The aim of this study was to determine the use of ANNs for the prediction of 2 subgroups of temporomandibular joint (TMJ) internal derangements (IDs) and normal joints using characteristic clinical signs and symptoms of the diseases. Materials and Methods: Clinical symptoms and diagnoses of 161 patients with TMJ ID were considered the gold standard and were employed to train a neural network. After the training process, the symptoms and diagnoses of 58 new patients were used to verify the network’s ability to diagnose. The diagnoses obtained from ANNs were compared with diagnoses of a surgeon experienced in temporomandibular disorders. The sensitivity and specificity of ANNs in predicting subtypes of TMJ ID were evaluated using clinical diagnosis as the gold standard. Results: Eight cases evaluated as bilaterally normal in clinical examination were evaluated as normal by ANN. In detecting unilateral anterior disc displacement with reduction (ADDwR; clicking), the sensitivity and specificity of ANN were 80% and 95%, respectively. In detecting unilateral anterior disc displacement without reduction (ADDwoR; locking), the sensitivity and specificity of ANN were 69% and 91%, respectively. In detecting bilateral ADDwoR, the sensitivity and specificity of ANN were 37% and 100%, respectively. In detecting bilateral ADDwR, the sensitivity and specificity of ANN were 100% and 89%, respectively. In detecting cases of ADDwR at 1 side and ADDwoR at the other side, the sensitivity and specificity of ANN were 44% and 93%, respectively. Conclusion: The application of ANNs for diagnosis of subtypes of TMJ IDs may be a useful supportive diagnostic method, especially for dental practitioners. Further research, including advanced network models that use clinical data and radiographic images, is recommended. © 2012 American Association of Oral and Maxillofacial Surgeons J Oral Maxillofac Surg 70:51-59, 2012 Temporomandibular disorder (TMD) is a generic term for several clinical signs and symptoms involving masticatory muscles, the temporomandibular joint (TMJ), and associated structures.1 Internal derange-
ment (ID) is 1 of the most common types of TMD.2 The term denotes an abnormal positional relation of the articular disc to the mandibular condyle and the articular eminence.3 TMJ ID is divided into 2 sub-
Received from Ondokuz Mayis University, Samsun, Turkey. *Assistant Professor, Faculty of Dentistry, Department of Oral and Maxillofacial Surgery. †Assistant Professor, Faculty of Engineering, Department of Electrical and Electronic Engineering. ‡Assistant Professor, Faculty of Dentistry, Department of Oral and Maxillofacial Surgery. §Research Assistant, Faculty of Dentistry, Department of Oral and Maxillofacial Surgery. 储Assistant Professor, Faculty of Dentistry, Department of Oral and
Maxillofacial Surgery. ¶Assistant Professor, Faculty of Dentistry, Department of Prosthodontics. Address correspondence and reprint requests to Dr Bekcioglu: Ondokuz Mayis University, Faculty of Dentistry, Department of Oral and Maxillofacial Surgery, 55139, Kurupelit, Samsun, Turkey; e-mail:
[email protected] © 2012 American Association of Oral and Maxillofacial Surgeons
0278-2391/12/7001-0$36.00/0 doi:10.1016/j.joms.2011.03.069
51
52 groups: anterior disc displacement with reduction (ADDwR; clicking) and anterior disc displacement without reduction (ADDwoR; locking).4 In ADDwR the intra-articular disc has slipped forward and mouth opening is accompanied by a clicking or popping sound. This percussive sound is produced as the condyle passes over the posterior band and returns to a normal relation with the disc. The clicking or popping noise generated during the translation phase is the most characteristic symptom of ADDwR. Patients with ADDwR may also be asymptomatic or they may have associated symptoms of joint pain and limitation of mouth opening. In ADDwoR the intra-articular disc is farther forward, and the condyle is unable to pass over the posterior band with attempted mouth opening. This condition is characterized by painful limitation of mouth opening, restricted lateral movements, and deviation to the affected side.5 After an acute episode, which can be extremely painful, the long-term condition can become nonpainful, with the range of motion approximating normal.1 Joint noise, limitation in mouth opening, joint pain, and mandibular deviation are the characteristic parameters in diagnosing TMJ IDs.6,7 Evaluation of a patient’s history and clinical examination findings has been the gold standard for diagnosing TMD.8-10 However, others have reported that TMJ abnormalities cannot be reliably assessed by clinical examination.11 The current clinical and research literature has suggested that the gold standard for the
ARTIFICIAL NEURAL NETWORK DIFFERENTIATION
diagnosis of TMJ disc displacement is an expert examiners’ decision based on all clinical and imaging data.12 Biologically inspired artificial neural networks (ANNs) are highly robust multifactorial mathematic models that have been applied successfully in the prediction, classification, function estimation, pattern recognition, and completion problems in many disciplines, including medicine.13 The first computational, trainable neural networks were developed in 1959.14,15 Since then, it has been used successfully in different medical applications.16 The use of ANNs in oral and maxillofacial surgery is limited, in some reports. There is a lack of literature on the use of ANNs in the differentiation of subtypes of TMJ ID. The purpose of this study was to use ANNs for the prediction of 2 common subgroups of TMJ ID and normal joints using characteristic clinical signs and symptoms of the diseases. The diagnoses obtained from ANN were compared with diagnoses of a specialist surgeon in TMD.
Materials and Methods Data were collected from 219 patients seeking TMD treatment at Ondokuz Mayis University, Faculty of Dentistry, Department of Oral and Maxillofacial Surgery. This study was exempt from institutional review board policy, as the data was collected in such
FIGURE 1. A feed-forward, back-propagation artificial neural network structure. DevL, deviation to left side; DevR, deviation to right side; Llat, left lateral movement; MaX, maximal mouth opening; Purelin, pure linear value; RLat, right lateral movement; SoundL, presence or absence of clicking in left joint; SoundR, presence or absence of clicking in right joint; Tanh, tangent sigmoid; VasL, visual analog scale value of left temporomandibular joint; VasR, visual analog scale value of right temporomandibular joint. Bas et al. Artificial Neural Network Differentiation. J Oral Maxillofac Surg 2012.
53
BAS ET AL
Table 1. DIAGNOSES (TARGET VALUES) PRODUCED AT OUTPUT LAYER
Output Layer (Targets)
Right
Left
1 2 3 4 5 6 7 8 9
N N N ADR ADR ADR ADNR ADNR ADNR
N ADR ADNR N ADR ADNR N ADR ADNR
Abbreviations: ADNR, anterior displacement without reduction; ADR, anterior displacement with reduction; N, normal joint. Bas et al. Artificial Neural Network Differentiation. J Oral Maxillofac Surg 2012.
a manner that subjects cannot be identified directly or indirectly. One hundred sixty-one patients with TMD were evaluated by an experienced oral and maxillofacial surgeon, and a final diagnosis was obtained by considering patients’ histories and clinical symptoms. All clinical diagnoses were made according to the Research Diagnostic Criteria of Temporomandibular Disorder guidelines. The clinical diagnosis was based on the following clinical signs: 1. Clicking: opening and closing clicks were carefully examined by bilateral digital palpation 2. Maximal mouth opening: measured intrinsically with a millimeter ruler with additional force produced by the examiner in an attempt to exclude the influence of masticatory muscle pain on mouth opening 3. Lateral movements: horizontal distance between the midpoints of the upper and lower incisors during left and right excursions 4. Deviation of the mandible: distance between the midpoints of the upper and lower incisors during maximal mouth opening
5. TMJ pain at the preauricular region during mandibular movements: patients were asked to rate their pain on a visual analog scale (0 ⫽ no pain; 10 ⫽ severe pain) The right and left TMJs of each patient were diagnosed as normal, ADDwR, or ADDwoR. Cases were clinically diagnosed and categorized: 1. 2. 3. 4. 5. 6. 7. 8. 9.
Bilateral normal joints Normal right TMJ and ADDwR in left TMJ Normal right TMJ and ADDwoR in left TMJ ADDwR in right TMJ and normal left TMJ Bilateral ADDwR ADDwR in right TMJ and ADDwoR in left TMJ ADDwoR in right TMJ and normal left TMJ ADDwoR in right TMJ and ADDwR in left TMJ Bilateral ADDwoR
The clinical symptoms and diagnoses of the first 161 patients were considered the gold standard and were used in training a neural network. After the training process, the symptoms and diagnoses of 58 new patients were used to verify the network’s ability to diagnose. DECISION-MAKING ALGORITHM
ANNs are widely used to approximate complex systems that are difficult to model using conventional modeling techniques such as mathematical modeling. The most common applications are function approximation (feature extraction), pattern recognition, and pattern classification. There is no exact available formula to decide which ANN architecture and which training algorithm will solve a given problem. The best solution is obtained by trial and error. One can obtain an idea by looking at a problem, starting with simple networks, and continuing to complex networks until the solution is within the acceptable limits of error. In this study, the decision-making algorithm was based on a back-propagation neural net-
Table 2. CALCULATED WEIGHT MATRIX AND BIAS VECTOR AT INPUT LAYER
Weight matrix (9 ⫻9) 0.0323 ⫺0.0042 ⫺0.0393 ⫺0.0245 ⫺0.0607 ⫺0.0779 0.0955 ⫺0.0665 0.0728 Bias vector (9 ⫻ 1) ⫺2.9321
⫺21.7977 0.1843 0.0838 ⫺0.1261 ⫺0.0340 2.6438 ⫺0.3686 0.5681 ⫺0.4383
0.0903 18.5054 1.0540 0.4396 0.2142 ⫺0.5347 0.0336 ⫺0.0825 ⫺0.7212
0.1099 ⫺0.0769 0.0825 0.1322 0.0750 ⫺0.0589 0.1133 0.0489 0.3138
0.1099 ⫺0.0147 ⫺0.1173 0.0349 0.0675 ⫺0.0164 0.0529 ⫺0.1703 0.1614
⫺23.6829 ⫺4.3782 ⫺8.8308 ⫺23.5734 ⫺19.7084 1.2926 3.9616 ⫺23.1705 ⫺5.3566
0.7019 23.2990 ⫺0.1422 1.2350 0.8239 4.0113 ⫺2.2934 ⫺22.9198 ⫺4.8879
⫺0.2020 17.8530 0.1723 ⫺0.0555 ⫺18.8442 24.2795 ⫺2.5936 ⫺0.5102 2.1719
⫺0.1996 ⫺1.4361 ⫺0.5267 ⫺0.5679 0.6091 0.3806 0.1653 ⫺19.3170 19.7707
4.3567
0.0323
⫺2.3099
⫺2.8630
0.7551
⫺2.8260
1.0262
⫺0.5359
Bas et al. Artificial Neural Network Differentiation. J Oral Maxillofac Surg 2012.
54
Table 3. CALCULATED WEIGHT MATRIX AND BIAS VECTOR AT HIDDEN LAYER Weight matrix (18 ⫻ 9) ⫺1.5261 ⫺3.6687 1.7164 0.4573 4.7052 ⫺1.1216 ⫺23.0422 ⫺0.1143 ⫺6.0122 Bias vector (18 ⫻ 1) 1.9705
0.6386 ⫺1.2086 ⫺0.7295 ⫺1.4032 ⫺0.6884 ⫺0.2951 0.5578 ⫺0.7499 ⫺0.5785
1.2322 ⫺5.6725 2.3298 ⫺0.8056 3.6358 0.8375 6.9282 0.8676 0.2078
2.7233 ⫺1.1564 ⫺5.1530 6.9734 9.1369 ⫺0.4835 ⫺3.3811 3.0310 ⫺8.9573
⫺6.6475 ⫺9.2258 6.5666 0.6531 22.5214 1.5412 ⫺9.3993 0.2025 ⫺19.5045
5.4761 9.6657 2.0880 ⫺0.0516 ⫺0.0913 ⫺1.8880 4.6294 ⫺0.4438 ⫺4.2869
0.5783 ⫺1.5491 3.1179 ⫺0.5820 1.3322 ⫺1.2018 11.4741 ⫺1.3598 ⫺1.0446
⫺2.2504 ⫺0.3303 ⫺0.0524 6.9766 1.0969 1.0752 ⫺0.1863 1.2377 ⫺1.2530
⫺0.6589 0.7878 ⫺0.9028 3.8338 0.3829 2.4348 2.2984 1.0452 1.1010
0.6993 ⫺1.8638 ⫺0.6859 1.4407 ⫺0.9549 ⫺2.8845 ⫺2.2194 2.7376 ⫺10.4928
⫺1.9918 ⫺1.1205 ⫺0.3850 ⫺0.0304 ⫺2.2011 0.9274 2.0703 ⫺0.4724 3.1766
⫺0.8428 ⫺1.8462 1.0811 2.8761 ⫺32.0240 ⫺1.3100 ⫺9.9875 ⫺1.1402 3.4989
3.8212 1.2315 2.7303 ⫺4.6651 ⫺1.1558 0.4545 ⫺0.6076 ⫺8.6225 ⫺14.3678
⫺3.8080 0.5304 2.2164 ⫺2.6298 ⫺0.8827 2.1717 3.1560 ⫺2.8412 6.3898
⫺2.7329 0.7175 ⫺2.8584 1.9410 0.4263 2.1862 ⫺1.5020 1.1575 1.5114
⫺0.2538 ⫺1.0452 0.9754 ⫺9.7002 11.1090 ⫺0.3289 0.7888 ⫺2.3006 2.9267
1.1744 1.0301 ⫺0.5663 2.5575 3.9368 0.1018 ⫺0.5605 0.1086 3.2806
⫺0.6781 0.6273 ⫺0.4376 2.1805 0.6027 2.3404 ⫺0.7826 ⫺1.7100 ⫺3.0094
⫺2.1567
1.6393
0.9338
⫺2.6442
1.1957
⫺2.9659
0.0597
⫺5.5829
⫺0.1143
⫺1.3597
⫺0.5096
⫺0.6690
⫺0.5922
1.7391
1.5636
2.1493
⫺1.2939
Bas et al. Artificial Neural Network Differentiation. J Oral Maxillofac Surg 2012.
Weight vector (18 ⫻ 8) ⫺1.4834 Bias vector (1 ⫻ 1) ⫺0.0180
⫺0.9472 1.9825 ⫺0.9559 ⫺1.1783 3.3052 ⫺3.0057 1.4998 ⫺0.3936 ⫺3.5890 ⫺1.1237 1.2287 0.5456 4.1810 2.0796 1.0593 0.3171 0.9541
Bas et al. Artificial Neural Network Differentiation. J Oral Maxillofac Surg 2012.
ARTIFICIAL NEURAL NETWORK DIFFERENTIATION
Table 4. CALCULATED WEIGHT MATRIX AND BIAS VECTOR AT OUTPUT LAYER
55
BAS ET AL
Table 5. NETWORK PARAMETERS USED IN DIAGNOSTICS
Parameters
Values
Learning coefficient Total iteration number Performance goal Activation neuron at input layer
0.8 20,000 10⫺4 Tansig
Abbreviation: Tansig, tangent sigmoid at hidden layer. Bas et al. Artificial Neural Network Differentiation. J Oral Maxillofac Surg 2012.
work. A back-propagation neural network is trained to perform complex functions in various fields of application, including pattern recognition, identification, classification, speech, vision, and control systems. The most widely used neural network is backpropagation. Back-propagation attempts to minimize error by adjusting each value of a network proportional to the derivative of error with respect to that value. This is called gradient descent. In back-propagation learning, the actual outputs are compared with the target values to derive the error signals, which are propagated backward layer by layer for the updating of synaptic weights in all lower layers. One of the most critical problems in constructing the ANN is the choice of the number of hidden layers and the number of neurons for each layer. Using few neurons in the hidden layer may prevent the training process to converge, whereas using too many neurons would produce a long training time and/or result in the ANN losing its generalization attribute. In this study, several tests were performed changing the number (1 or 2) of hidden layers and changing the number of neurons in each hidden layer with full connections between neurons. As shown in Figure 1,
an ANN structure with an input layer including 9 neurons, a hidden layer including 18 neurons, and an output layer including 1 neuron is assumed adequate for classifying 9 inputs. This architecture proved capable of approximating any function with a finite number of discontinuities with arbitrary accuracy. It was used to discriminate 9 decisions at the output layer. Table 1 presents the decisions (diagnoses). The abbreviations used in the report are given in nomenclature. The input vectors consisted of 9 variables, namely maximal mouth opening (centimeters), visual analog scale value of right TMJ, visual analog scale value of left TMJ, right lateral movement (centimeters), left lateral movement (centimeters), presence or absence of clicking in the right joint (1 ⫽ clicking; 0 ⫽ absence of clicking), presence or absence of clicking in the left joint (1 ⫽ clicking; 0 ⫽ absence of clicking), deviation to the right side (centimeters), and deviation to the left side (centimeters). Data from 161 consecutive patients formed a 161 ⫻ 9 matrix for the training procedure. After the training procedure, the calculated weight matrix and bias vector at the input layer (Table 2), hidden layer (Table 3), and output layer (Table 4) were calculated. The activation neurons were selected for the tangent sigmoid at the input and hidden layers and for the pure linear value at the output layer. The training algorithm was chosen for resilient back-propagation. Table 5 presents the network parameters. Performance of the proposed ANN for diagnostics was tested using data from 58 different consecutive patients. The entire calculation process was developed using Visual Basic 6.0 (Microsoft, Redmond, WA). The Start button was the main function that computed all required calculations according to pa-
FIGURE 2. Working window of Start button with a target value. ADNR, anterior displacement without reduction; DEV L, deviation to left side; DEV R, deviation to right side; L LAT, left lateral movement; MAX, maximal mouth opening; R LAT, right lateral movement; SOUND L, presence or absence of clicking in left joint; SOUND R, presence or absence of clicking in right joint; VAS L, visual analog scale value of left temporomandibular joint; VAS R, visual analog scale value of right temporomandibular joint. Bas et al. Artificial Neural Network Differentiation. J Oral Maxillofac Surg 2012.
56
ARTIFICIAL NEURAL NETWORK DIFFERENTIATION
FIGURE 3. Working window of Start button without a target value. ADNR, anterior displacement without reduction; C, current integer value of predicted value (diagnosis); DEV L, deviation to left side; DEV R, deviation to right side; L, ⫺1 of integer value of predicted value; L LAT, left lateral movement; MAX, maximal mouth opening; R LAT, right lateral movement; SOUND L, presence or absence of clicking in left joint; SOUND R, presence or absence of clicking in right joint; VAS L, visual analog scale value of left temporomandibular joint; VAS R, visual analog scale value of right temporomandibular joint; U, rounded-up value of integer value of predicted value. Bas et al. Artificial Neural Network Differentiation. J Oral Maxillofac Surg 2012.
tients’ test data. Figure 2 shows the working window of the Start button. The user can easily input patients’ test data into the required areas. If the target value is known, the ANN simply predicts the diagnosis and decides the diagnostics by pressing the Predict button. Then, the Statistics button calculates the overall error and performance of the ANN. In this particular example, the calculated diagnostic value is 9.0060 and overall error is 0.07%. The performance of the ANN for this particular example is 99.933%. If the target is unknown, the statistical calculations are performed according to the predicted value. Figure 3 shows the working window of the Start button without a target value. In this case, 0 is entered in the target section. The integer value of the diagnosis (predicted value) is used for a 0 target value. In Figure 3, “L” denotes ⫺1 of the integer value of the predicted value. In this particular example, the calculated target value is assumed to be 8 and total error is calculated as 12.58%. “C” denotes the current integer value of the predicted value (diagnosis). In this case, the calculated target value is assumed to be 9 and total error is calculated as 0.07%. “U” denotes the rounded-up value of the integer value of the predicted value. The rounded-up value is 10, and the overall error is calculated as 9.94%. For this particular example, the diagnosis is ADDwoR in the 2 sides. The relative error presented in equation 1 is used to calculate the error between the actual and predicted target values. error 共%兲 ⫽ 100(xactual ⫺ xpredicted) ⁄ xactual
(1)
The user can import the patients’ data file from the File tab and save the result for after recording. The performance of the suggested ANN structure for med-
ical diagnostics was tested using data from 58 consecutive patients.
Results The network model was first trained using data pertaining to clinical symptoms and diagnoses of 161 patients. Table 6 presents the number of patients whose right and left TMJs were clinically diagnosed as normal, ADDwR, or ADDwoR. Fifty-eight new cases were used to verify the network’s ability to diagnose. Table 7 presents the number of new patients whose right and left TMJs were clinically diagnosed as normal, ADDwR, or ADDwoR. Table 8 presents the 58
Table 6. PATIENTS IN TRAINING GROUP WHOSE RIGHT AND LEFT TEMPOROMANDIBULAR JOINTS WERE CLINICALLY DIAGNOSED AS NORMAL OR ANTERIOR DISC DISPLACEMENT WITH OR WITHOUT REDUCTION
Right TMJ
Left TMJ
Number of Patients
N N N ADDwR ADDwR ADDwR ADDwoR ADDwoR ADDwoR
N ADDwR ADDwoR N ADDwR ADDwoR N ADDwR ADDwoR
12 9 38 9 20 12 29 13 19
Abbreviations: ADDwoR, unilateral anterior disc displacement without reduction (locking); ADDwR, unilateral anterior disc displacement with reduction (clicking); N, normal; TMJ, temporomandibular joint. Bas et al. Artificial Neural Network Differentiation. J Oral Maxillofac Surg 2012.
57
BAS ET AL
Table 7. NUMBER OF PATIENTS IN TEST GROUP WHOSE RIGHT AND LEFT TEMPOROMANDIBULAR JOINTS WERE CLINICALLY DIAGNOSED AS NORMAL OR ANTERIOR DISC DISPLACEMENT WITH OR WITHOUT REDUCTION
Right TMJ
Left TMJ
Number of Patients
N N N ADDwR ADDwR ADDwR ADDwoR ADDwoR ADDwoR
N ADDwR ADDwoR N ADDwR ADDwoR N ADDwR ADDwoR
8 5 7 5 10 5 6 4 8
Abbreviations: ADDwoR, unilateral anterior disc displacement without reduction (locking); ADDwR, unilateral anterior disc displacement with reduction (clicking); N, normal joints; TMJ, temporomandibular joint. Bas et al. Artificial Neural Network Differentiation. J Oral Maxillofac Surg 2012.
new patients whose right and left TMJs were predicted as normal, ADDwR, or ADDwoR by the ANN. Figure 4 shows the relative errors between the actual and predicted target values grouped according to Table 1. The relative error percentages were higher than 10% in only 15 ANN predictions. The overall performance of the proposed diagnosis system is presented in Table 9, which shows the average relative error as a percentage for each diagnosis (ANN predictions) in descending order. The sensitivity and specificity of ANNs in predicting subtypes of TMD were evaluated using clinical diagnosis as the gold standard. Eight cases evaluated
Table 8. NUMBER OF PATIENTS IN TEST GROUP WHOSE RIGHT AND LEFT TEMPOROMANDIBULAR JOINTS WERE PREDICTED AS NORMAL OR ANTERIOR DISC DISPLACEMENT WITH OR WITHOUT REDUCTION BY THE ARTIFICIAL NEURAL NETWORK
Right TMJ
Left TMJ
Number of Patients
N N N ADDwR ADDwR ADDwR ADDwoR ADDwoR ADDwoR
N ADDwR ADDwoR N ADDwR ADDwoR N ADDwR ADDwoR
8 5 6 5 15 3 7 5 4
Abbreviations: ADDwoR, unilateral anterior disc displacement without reduction (locking); ADDwR, unilateral anterior disc displacement with reduction (clicking); N, normal; TMJ, temporomandibular joint. Bas et al. Artificial Neural Network Differentiation. J Oral Maxillofac Surg 2012.
as bilaterally normal in clinical examination were evaluated as normal by the ANN. Ten cases were diagnosed as unilateral ADDwR in clinical examination; 2 were misdiagnosed by the ANN (80% sensitivity, 95% specificity). Thirteen cases were diagnosed as unilateral ADDwoR in clinical examination; 4 were misdiagnosed by the ANN (69% sensitivity, 91% specificity). Ten cases were diagnosed as bilateral ADDwR in clinical examination; 0 were misdiagnosed by the ANN (100% sensitivity, 89% specificity). Eight cases were diagnosed as bilateral ADDwoR in clinical examination; 5 were misdiagnosed by the ANN (37% sensitivity, 100% specificity). Nine cases were diagnosed as ADDwR in 1 side and ADDwoR in the other side during clinical examination; 5 were misdiagnosed by the ANN (44% sensitivity, 93% specificity).
Discussion TMD encompasses several overlapping conditions.17 An average of 40% to 60% of the population has at least 1 TMD symptom and an average of 3.6% to 7% has TMD with sufficient severity that treatment is desired.18 The diagnosis and treatment of patients with TMD is a challenge for dental practitioners, especially those who have not trained in maxillofacial surgery. The diagnosis of TMD is usually performed by clinical investigation combined with imaging techniques.19 The use of imaging techniques in diagnosing TMD is still controversial. One group has recommended that practitioners should use imaging only if there is a reasonable expectation that additional information will influence a patient’s treatment approach.20 Different ANN structures are valuable models used in the medical field for the development of decision support systems. The use of ANNs in oral and maxillofacial surgery has been limited. Speight et al21 evaluated the use of a neural network to identify patients at risk of oral cancer and precancer and reported successful results in identifying patients with a high risk of oral cancer. Brickley et al22 developed and tested 12 neural networks of different architectures for lower third molar treatment planning decisions and reported high sensitivity and specificity compared with decisions by a senior oral surgeon. In another report, Brickley and Shepherd23 suggested that a computer-based neural network could play a useful role in supporting dental practitioners in making third molar referral decisions. Park et al24 evaluated cervical lymph node metastasis of squamous cell carcinoma using a neural network and compared it with magnetic resonance imaging criteria. Radke et al25 used an ANN for differentiation of normal TMJs from ADDwoR that analyzed average frontal chewing patterns. These studies found positive results in using
58
ARTIFICIAL NEURAL NETWORK DIFFERENTIATION
FIGURE 4. Relative errors between actual and predicted target values grouped according to Table 1. Bas et al. Artificial Neural Network Differentiation. J Oral Maxillofac Surg 2012.
neural networks for dental diagnosis. A neural network therefore can be developed to assist dentists in achieving correct interpretations and decreasing human error. To the best of our knowledge, this is the first study to use ANNs as a supportive diagnostic method in detecting subtypes of TMJ ID. Neural network techniques should classify a disease if the features of a disease pattern are well defined and can be quantified as values or vectors.26 In this study, 1 major shortcoming of the technique was its insufficiency in detecting disc displacements that do not show the characteristic symptoms of a disease. Bilateral ADDwoR showed the highest average relative error in this report, whereas bilateral ADDwR showed the lowest relative error. In our opinion, this was the result of the different symptoms in ADDwoR. In contrast, ADDwR has well-defined symptoms that minimize misdiagnosis. ADDwoR can be present in acute
Table 9. AVERAGE RELATIVE ERRORS OF EACH DIAGNOSIS
Diagnosis
Relative Error (%)
Bilateral ADDwoR Right ADDwR, left ADDwoR Right N, left ADDwoR Right N, left ADDwR Right N, left N Right ADDwoR, left ADDwR Right ADDwR, left N Right ADDwoR, left N Right ADDwR, left ADDwR
24.22 13.01 10.97 10.46 8.43 8.28 7.85 5.43 0.03
Abbreviations: ADDwoR, unilateral anterior disc displacement without reduction (locking); ADDwR, unilateral anterior disc displacement with reduction (clicking); N, normal. Bas et al. Artificial Neural Network Differentiation. J Oral Maxillofac Surg 2012.
or chronic form, each presenting different clinical symptoms. In our opinion, more accurate results can be obtained when the network model is trained with clinical symptoms and magnetic resonance images of patients. In addition, when more data are fed to the training set of the ANN, the normalized system error can be decreased. In conclusion, diagnosis of TMDs is a challenge for dental practitioners, especially those who have not trained in maxillofacial surgery. ANNs could be developed to assist dentists in achieving correct interpretations and decreasing human error. This was a preliminary study in the application of ANNs for diagnosis of subtypes of TMJ ID. More accurate predictions could have been obtained if the ANN has been trained with additional data from more patients. The proposed ANN-based diagnostic system is easy to use and developed in Visual Basic 6.00. It can be a useful decision support method in diagnosing subtypes of TMD, especially for dental practitioners. Further research is recommended, including research on advanced network models that use clinical and imaging data.
References 1. McNeill C: Management of temporomandibular disorders: Concepts and controversies. J Prosthet Dent 77:510, 1997 2. Wright EF: Manual of Temporomandibular Disorders. Ames, IA, Blackwell Publishing, 2005, p 61 3. Brandlmaier I, Rudisch A, Bodner G, et al: Temporomandibular joint internal derangement: Detection with 12.5 MHz ultrasonography. J Oral Rehabil 30:796, 2003 4. Foster ME, Gray RJ, Davies SJ, et al: Therapeutic manipulation of the temporomandibular joint. Br J Oral Maxillofac Surg 38:641, 2000 5. Laskin DM, Grene CS, Hylander WL: Temporomandibular Disorders. An Evidence-Based Approach to Diagnosis and Treatment. Hanover Park, IL, Quintessence Publishing, 2006, p 128
BAS ET AL 6. Katzberg RW, Anderson QN, Helms CA: Arthrography, in Helms, CA, Katzberg, RW, Dolwick, MF (eds): Internal Derangements of the Temporomandibular Joint. San Francisco, Radiology Research and Education Foundation, 1983, p 85 7. Katzberg RW, Dolwick MF, Helms CA, et al: Arthrotomography of the temporomandibular joint. AJR Am J Roentgenol 134:995, 1980 8. Yatani H, Sonoyama W, Kuboki T, et al: The validity of clinical examination for diagnosing anterior disk displacement with reduction. Oral Surg Oral Med Oral Pathol Oral Radiol Endod 85:647, 1998 9. Truelove EL, Sommers EE, LeResche L, et al: Clinical diagnostic criteria for TMD: New classification permits multiple diagnosis. J Am Dent Assoc 123:47, 1992 10. Okeson JP: Current terminology and diagnostic classification schemes. Oral Surg Oral Med Oral Pathol Oral Radiol Endod 83:61, 1997 11. Tvrdy P: Methods of imaging in the diagnosis of temporomandibular joint disorders. Biomed Pap Med Fac Univ Palacky Olomouc Czech Repub 151:133, 2007 12. Orsini MG, Kuboki T, Terada S, et al: Clinical predictability of temporomandibular joint disc displacement. J Dent Res 78:650, 1999 13. Dayhoff JE, DeLeo JM: Artificial neural networks: Opening the black box. Cancer 91:1615, 2001 14. Widrow B, Hoff M: Adaptive switching circuits. August IRE WESCON Convention Rec 4:96, 1960 15. Widrow B, Stearns SD: Adaptive Signal Processing. Englewood Cliffs, NJ, Prentice-Hall, 1985 16. Reggia JA: Neural computation in medicine. Artif Intell Med 5:143, 1993 17. Ethunandan M, Wilson AW: Temporomandibular joint arthrocentesis—More questions than answers? J Oral Maxillofac Surg 64:952, 2006
59 18. Okeson JP: Management of Temporomandibular Disorders and Occlusion (ed 5). St Louis, Mosby, 1998, p 153 19. Larheim TA, Westesson PL: TMJ imaging, in Laskin DM, Grene CS, Hylander WL (eds): Temporomandibular Disorders. An EvidenceBased Approach to Diagnosis and Treatment. Hanover Park, IL, Quintessence Publishing, 2006, p 128 20. Tyndall DA, Brooks SL: Selection criteria for dental implant site imaging: A position paper of the American Academy of Oral and Maxillofacial Radiology. Oral Surg Oral Med Oral Pathol Oral Radiol Endod 89:531, 2000 21. Speight PM, Elliott J, Downer M, et al: The use of artificial intelligence to identify people at risk of oral cancer and precancer. Br Dent J 179:383, 1995 22. Brickley MR, Shepherd JP, Armstrong RA: Neural networks: A new technique for development of decision support systems in dentistry. J Dent 26:305, 1998 23. Brickley MR, Shepherd JP: Comparisons of the abilities of a neural network and three consultant oral surgeons to make decisions about third molar removal. Br Dent J 182:59, 1997 24. Park SW, Heo MS, Lee SS, et al: Artificial neural network system in evaluating cervical lymph node metastasis of squamous cell carcinoma. Korean J Oral Maxillofac Radiol 29:149, 1999 25. Radke JC, Ketcham R, Glassman B, et al: Artificial neural network learns to differentiate normal TMJs and nonreducing displaced disks after training on incisor-point chewing movements. Cranio 21:259, 2003 26. Lo Shih-Chung B, Lin-Jyh-Shyan J, Freedman-Matthew T, et al: Application of artificial neural networks to medical image pattern recognition: Detection of clustered microcalcifications on mammograms and lung cancer on chest radiograph. J VLSI Signal Process Syst Signal Image Video Technol 18:263, 1998
