- Email: [email protected]
The Reliability of Artificial Neural Network in Locating Minor Apical Foramen: A Cadaver Study
Basic Research—Technology
The Reliability of Artificial Neural Network in Locating Minor Apical Foramen: A Cadaver Study Mohammad Ali Saghiri, BSc, MSc, PhD,* Franklin Garcia-Godoy, DDS, MS, PhD,† James L. Gutmann, DDS, PhD,‡ Mehrdad Lotfi, DMD, MSc,§ and Kamal Asgar, BSc, MSc, PhDjj Abstract Introduction: The purpose of this study was to evaluate the accuracy of the artificial neural network (ANN) in a human cadaver model in an attempt to simulate the clinical situation of working length determination. Methods: Fifty single-rooted teeth were selected from 19 male cadavers ranging in age from 49–73 years. Access cavities were prepared, a file was placed in the canals, and the working length was confirmed radiographically by endodontists. The location of the file in relation to the minor apical foramen was categorized as long, short, and exact by the ANN, by endodontists before extraction, and stereomicroscopically after extraction. The results were compared by using Friedman and Wilcoxon tests. The significance level was set at P <.05. Results: The Friedman test revealed a significant difference among groups (P < .001). There were significant differences between data obtained from endodontists and ANN (P = .001) and data obtained from endodontists and real measurements by stereomicroscope after extraction (P < .002). The correct assessment by the endodontists was accurate in 76% of the teeth. ANN determined the anatomic position correctly 96% of the time. The confidence interval for the correct result was 64.16–87.84 for endodontists and 90.57–101.43 for ANN. Conclusions: ANN was more accurate than endodontists’ determinations when compared with real working length measurements by using the stereomicroscope as a gold standard after tooth extraction. The artificial neural network is an accurate method for determining the working length. (J Endod 2012;38:1130–1134)
Key Words Apical foramen, artificial neural network, cadaver, root canal treatment, working length
C
orrect working length (WL) determination is an essential step in achieving successful root canal treatment outcomes. Instrumentation beyond the apical foramen (1), flare-ups (2), periapical foreign body reactions (3), and poor microbial control (4) are the result of poor WL determination and control. In 1930 Grove (5) stated that the proper point to which root canals should be cleaned and filled is the junction of the dentin and the cementum and that the pulp should be severed at the point of its union with the periodontal membrane. The cementodentinal junction (CDJ) is the anatomic and histologic landmark where the periodontal ligament begins and the pulp ends. Although the CDJ is a practical termination point for the preparation and obturation of the root canal, this anatomic point cannot be determined radiographically. However, for decades academic programs taught that the root canal procedures of shaping, enlarging, and obturation should be at or slightly short of the apical location. At the First World Endodontic Conference held in Philadelphia in 1953, 2 of the universal principles agreed upon were the following: (1) Traumatic injury to the surrounding (periapical) soft tissue should be avoided at all times. To this end, instrumentation stops should be used, and instruments should be confined entirely with the root canal. (2) The canal filling should seal the apical foramen, and if the apical millimeter or so of the canal is filled with healthy living tissue, the root canal filling should terminate at this level rather than at the apical foramen (6). An in vivo histologic study found that the most favorable histologic conditions were when the instrumentation and obturation remained short of the apical constriction and that extruded gutta-percha and sealer always caused a severe inflammatory reaction despite the absence of pain (7). The ability to achieve this position predictably is based on establishing a sound WL for each root. There are various methods for locating the apical foramen and determining the WL, including radiographic, digital tactile sense, and patients’ response to a file or paper point (8). Although little documentation exists to support the digital and patients’ response techniques (9), radiography is an accepted, commonly used method (10, 11). Some original radiographic methods requiring formulas, such as those of Bramante and Berbert (12), have been largely abandoned. Electronic apex locators (EALs) are very useful adjuncts during root canal procedures for locating the apical foramen (13), and recently, cone-beam computed tomography has also been determined to be an accurate means for determining the WL (14). The principles behind the first electronic device for root length determination were presented by Custer (15) in 1918 and commercialized by Sunada (16) in 1962 by using a device that registered consistent values in electrical resistance between an instrument in a root canal and an electrode on the oral mucous membrane (17). When reviewing the vast amount of literature that has addressed the efficacy and
From the *Department of Dental Material, Islamic Azad University (Tehran Branch), Tehran, Iran; †Bioscience Research Center, College of Dentistry, University of Tennessee Health Science Center, Memphis, Tennessee; ‡Department of Restorative Sciences, Baylor College of Dentistry, Texas A&M Health Science Center, Dallas, Texas; §Research Center for Pharmaceutical Nanotechnology and Department of Endodontics, Dental Faculty, Tabriz University (Medical Sciences), Tabriz, Iran; and k Department of Dental and Biological Materials, University of Michigan, Ann Arbor, Michigan. Address requests for reprints to Dr Mohammad Ali Saghiri, Assistant Professor, Department of Dental Material, Kamal Asgar Research Center (KARC) and Dental School, Azad University (Tehran Branch), P.O. Box 14665-1445, Tehran, Iran. E-mail address: [email protected] 0099-2399/$ - see front matter Copyright ª 2012 American Association of Endodontists. doi:10.1016/j.joen.2012.05.004
1130
Saghiri et al.
JOE — Volume 38, Number 8, August 2012
Basic Research—Technology accuracy of EALs, some in vitro experimental models give greater accuracy than can be achieved clinically (13, 18). No individual technique, however, is truly satisfactory in determining root canal WL (11). Digital radiography used for WL determination has the advantages of lower radiation exposure to the patient, ability to manipulate the digitized image for better observation, ease of radiographic data storage, and instant display of the image (19). The quality of the image is probably of utmost significance in endodontics because it facilitates accurate interpretation of root and canal morphology (20). However, clinical errors in radiographic interpretation by using standard films (21) and digital radiographs (22) create a need for an objective, computer-based analysis of WL. An artificial neural network (ANN) is a mathematical computerbased model inspired by the structure and/or functional aspects of biological neural networks in the brain. ANN is a decision-making system and helps the diagnostic procedure used for prediction of different elements from radiographs. ANNs are computer models with a massive parallel structure that imitate the human brain (23). They consist of nonlinear calculation elements, which represent neurons organized in several highly interconnected layers inspired by biological neural networks (24). The basic idea is that even though artificial neurons have no intelligence individually, when interconnected, they might be able to duplicate aspects of the human brain combined with the computational power of computers. ANNs have proved to be extremely helpful in solving problems with high computational complexity. ANNs have also been established as promising alternatives to statistical discriminate analysis because they can synthesize a considerable amount of information (or variables) without requiring statistical modeling of the problem (25). Because medical diagnosis is one of typical pattern recognition problem, previous applications have inspired the use of ANNs in the dentistry field such as cephalometric diagnosis (26) and differentiation of subgroups of temporomandibular internal derangements (27). Recently an investigation (28) showed that ANNs can act as a second opinion to locate the apical foreman on radiographs to enhance the accuracy of WL determination by radiography and can function as a decision-making system in various similar clinical situations. The lack of objectiveness in determination of WL coupled with high levels of variability among clinicians about the location of the file on the radiograph can impact greatly on the treatment rendered, patient discomfort, and successful outcomes. Therefore, the aim of this study was to evaluate the accuracy of an objective method of determining the WL in a human cadaver model and compare it with endodontists’ interpretations and microscopic realities.
Materials and Methods The research protocol was approved by the Research Ethics Committee of Kamal Asgar Research Center (protocol no. KARC/ 10B2011-70-20). Fifty single-rooted teeth with curvatures <30 in the apical region were selected from fresh cadavers that were males ranging in age between 49 and 73 years. This part of the study was similar to our previous study (28). Briefly, access cavities were prepared, and WLs were determined by placing a #20 K-file (Dentsply Maillefer, Ballaigues, Switzerland) in the root canal and exposing radiographs by using a Rinn XCP (Dentsply Rinn, Elgin, IL) and a photostimulable phosphorus system (Digora, Soredex, Tuusula, Finland). The files were fixed in the tooth by using cyanoacrylate glue (Pattex; Henkel AG & Co, Duesseldorf, Germany). The teeth were extracted, and after a gentle wash with normal saline the apical portions of the teeth were assessed by using a stereomicroscope (Nikon Corporation, JOE — Volume 38, Number 8, August 2012
Tokyo, Japan) equipped with a CCD camera at 20 to determine the position of the files. The positions were classified as follows: (1) File tip was not visible under the stereomicroscope (short). (2) The file tip was visible under the stereomicroscope, but it was not visible more than 0.3 mm beyond the minor foramen (exact). (3) The file was visible under the stereomicroscope and was more than 0.3 mm beyond the minor foramen (long). The prototype based on multilayer perceptron ANN model in combination with the Otsu method and wave let theory that had been used in previous study (28) was used to determine the WL (Fig. 1). Two endodontists evaluated the position of the files according to the position of the apical foramen on the radiographs, which visualized the root apices under radiographs by using a Rinn XCP and a photostimulable phosphorus system. If there were different opinions regarding the position of the file, another endodontist evaluated the radiographs. At least 2 similar opinions were considered as the final decision. The results of the endodontists’ evaluation, the ANN, and stereomicroscope assessment were compared by using Friedman and Wilcoxon tests. The significance level was set at P <.05.
Results The Friedman test revealed a significant difference among groups (P < .001). The Wilcoxon test was used to reveal any significant differences between specific groups. There were significant differences between data obtained from endodontists and ANN (P = .001) and data obtained from the endodontists and real measurements obtained from the stereomicroscopic assessment (P < .002) (Fig. 2). However, there were no significant differences between measurements of root canal length when ANN was compared with real measurements after extraction (P = .74). The determination of the minor anatomic constriction correctly by endodontists occurred 76% of the time, whereas the ANN identified this position correctly 96% of the time. Confidence interval for correct results was 64.16–87.84 for endodontists and 90.57–101.43 for ANN.
Discussion In most teeth requiring root canal procedures, the apical foramen is located on the lateral surface of the root at an average distance of 0.36 mm for anterior teeth and 0.48 mm for posterior teeth. Martos et al (29) found that the average distance between the apical foramen and the anatomic apex was 0.69 mm. This distance, however, was different in posterior teeth (0.82 mm) when compared with anterior teeth (0.39 mm); thus, in 60% of cases, the apical foramen is located on a lateral surface of the root and in 40% of cases on the anatomic apex. The same authors illustrated that the lateral position of the apical foramen has been detected in 43% of posterior teeth and in 17% of anterior teeth (30). WL can be determined with radiographs, tactile sensation, and EALs. However, radiographs are subject to distortion, magnification, interpretation variability, and lack of 3-dimensional representation. As a result, WL is generally measured about 0.5–1 mm short of the radiographic apex. Pratten and McDonald (31) showed that considering the position of the apical constriction located 1 mm short of the radiographic apex will result in an underestimation of WL. Vertical and horizontal cone angulations, film position, tooth inclination, and film processing issues could also influence WL determination by using radiographs (21). In some cases, the apical foramen might be located on the lateral root surface up to 3.5 mm from the radiographic apex (32). In such teeth, if the canal terminates coronal to the anatomic
Artificial Neural Network in Endodontics
1131
Basic Research—Technology
Figure 1. (A) Sample of radiograph (exact) and (B) stereomicroscope (long) and (C) snapshot of ANN software.
apex and in the plane of the film, the radiographic appearance will be short, and any adjustment will result in an overestimation of the WL. Numerous studies have reported on the accuracy of EALs for determining the location of the apical foramen. EALs can accurately determine the WL in 75.0%–96.5% of the root canals with mature apices (13, 33). In the presence of conductive restorations, severely undermined caries, serous, purulent, or hemorrhagic exudates, wide-canal or a wide-open apex WL determination is not accurate (34). In the present study EAL methods of WL determination were not used for 3 reasons: first, the blood pressure of cadavers was not the same as in vivo studies (35); second, the temperatures of corpses were lower than clinical scenario, which may impact on results (36); and third, in the presence of cellular death, the electric potential of cells is vastly different from that in viable cells (37), which may influence conductivity. The validity of measurements made with an in vitro experimental protocol, that is, the extent to which they depict the clinical accuracy of EALs, is unknown (17). However, they are able to reproduce the clinical condition of EAL use and facilitate an objective examination of several variables that are not practical in in vivo studies (33). WL determination by using radiographs is indispensable part of root canal treatment (38). The use of a combination of methods to determine an accurate WL may be more successful than relying on just 1 method (39). A parallel radiograph seems to be more accurate for determining the location of the apical foramen in human teeth during root canal
instrumentation (40) and a better reproduction of the distance between the apex and a contrasting subject (eg, a root canal instrument) (41). However, it has been stated that the accurate determination or even estimation of the apical canal constriction is not possible with radiography because of anatomic variations or errors in projection (34, 42). A neural network is a combination of soft and hardware simulation of a biological brain. The aim of a neural network is to learn to recognize patterns from entered data. Once the neural network has been trained on entered data, it can make predictions by detecting similar patterns in future data. Neural networks are a branch of biomedical engineering known as artificial intelligence. Other branches include genetic algorithms, expert systems, fuzzy logic, and chaos theory (43). ANN can be considered as a black box that is able to predict an output pattern when it recognizes a given input pattern. This network must first be "trained" by having it process a large number of input patterns and showing it what output resulted from each input pattern. Once trained, the network is able to recognize similarities when presented with a new input pattern, resulting in a predicted output pattern (27). In the current study we used a prototype that trained in a previous study (28) and the Otsu feature as a built-in. The Otsu method was used to automatically perform shape-based image thresholding histogram. This method can separate the teeth from surrounding tissues on the basis of differences in gray scales on radiographs (28). The teeth and surrounding tissues have many distinct features such as color and intensity. The algorithm assumes that the image to be thresholded
Figure 2. The plot of validity of neural network (A) and endodontists (B) showed that the correct identification for endodontists was 76%, whereas ANN determination was correct 96% of the time.
1132
Saghiri et al.
JOE — Volume 38, Number 8, August 2012
Basic Research—Technology contains 2 classes of pixels. The white color of the teeth produces 1 pinnacle in the histogram, and gray scales lower than the peaks (threshold) are removed by first determining a threshold independently. Only differences between the white color of the teeth and surrounding tissues remain. The Otsu method searches for a threshold that minimizes the intraclass variance, defined as a weighed sum of variances of the 2 classes (44). The major advantage of ANNs is their ability to be used as an arbitrary function approximation mechanism that learns from observed data. It should be mentioned that ANN systems are close dependent to pool of data that are used for training, so if we increase the pool of data for training system, logically we should be more accurate (23). ANN is capable of learning important relationships from a set of data and applying this knowledge to evaluate new cases. Also it modifies its behavior (trains) by adjusting the strength or weights of the connections until its own output converges to the known correct output. Once trained, the network can evaluate a new case of input values by applying the weights learned from the data set with which it was trained (45). The network can generalize from previous cases to evaluate cases not previously seen. It has been stated that ANN can act as a second opinion for locating the apical foramen in an in vitro situation (28). The present study showed that the cadavers can be used as a model to examine questions regarding the soft tissue interferences with the function and application of the ANN in clinical situations.
Conclusion Within the limitations of the present experiments, the following could be concluded. ANN is an accurate method for determining the WL in human cadavers. Incorporation of surrounding tissue does not have a deleterious effect on ANN decision-making. ANN determined the anatomic position correctly 96% of the time. Observations of ANN results suggest that it might be considered as an alternative for EAL or even confirmation of the results obtained with an EAL; however, further investigation is recommended.
Acknowledgments This project is dedicated to Sophie Asgar (late wife of Professor Asgar), who died during experimental part of present study. In addition, we thank Professor Mohammad Aeinehchi for providing all facilities for working on cadavers, Professors Shahram Azimi and Payman Mehrvazfar for making score and contributing to analysis, and Dr Houtan Aghili from IBM, New York for his invaluable help and re-check of our network and program. The authors deny any conflicts of interest related to this study.
References 1. Smith CS, Setchell DJ, Harty FJ. Factors influencing the success of conventional root canal therapy: a five-year retrospective study. Int Endod J 1993;26:321–33. 2. Georgopoulou M, Anastassiadis P, Sykaras S. Pain after chemomechanical preparation. Int Endod J 1986;19:309–14. 3. Nair MK, Nair UP. Digital and advanced imaging in endodontics: a review. J Endod 2007;33:1–6. 4. Chugal NM, Clive JM, Spangberg LS. Endodontic infection: some biologic and treatment factors associated with outcome. Oral Surg Oral Med Oral Path Oral Radio Endod 2003;96:81–90. 5. Grove C. Why canals should be filled to the dentinocemental junction. J Am Dent Assoc 1930;17:293–6. 6. Grossman LI. Transactions of the World Conference on Endodontics. Philadelphia: University of Pennsylvania; 1953.
JOE — Volume 38, Number 8, August 2012
7. Ricucci D, Langeland K. Apical limit of root canal instrumentation and obturation, part 2: a histological study. Int Endod J 1998;31:394–409. 8. Seidberg B, Alibrandi B, Fine H, Logue B. Clinical investigation of measuring working lengths of root canals with an electronic device and with digital-tactile sense. J Am Dent Assoc 1975;90:379–87. 9. Gutmann JL, Lovdahl PE. Problem Solving in Endodontics: Prevention, Identification, and Management, 5th ed. Maryland Heights, MO: Elsevier/Mosby; 2011. 10. Powell-Cullingford AW, Pitt Ford TR. The use of E-speed film for root canal length determination. Int Endod J 1993;26:268–72. 11. Gutmann JL, Leonard JE. Problem solving in endodontic working-length determination. Compend Contin Educ Dent 1995;16:288. 290, 293–284 passim; quiz 304. 12. Bramante CM, Berbert A. A critical evaluation of some methods of determining tooth length. Oral Surg Oral Med Oral Pathol 1974;37:463–73. 13. Gordon MP, Chandler NP. Electronic apex locators. Int Endod J 2004;37:425–37. 14. Janner SF, Jeger FB, Lussi A, Bornstein MM. Precision of endodontic working length measurements: a pilot investigation comparing cone-beam computed tomography scanning with standard measurement techniques. J Endod 2011; 37:1046–51. 15. Custer L. Exact methods of locating the apical foramen. J Natl Dent Assoc 1918;5: 815–9. 16. Sunada I. New method for measuring the length of the root canal. J Dent Res 1962; 41:375–87. 17. Kim E, Lee SJ. Electronic apex locator. Dent Clin North Am 2004;48:35–54. 18. Venturi M, Breschi L. A comparison between two electronic apex locators: an ex vivo investigation. Int Endod J 2007;40:362–73. 19. Akcay I, Ilhan B, Dundar N. Comparison of conventional and digital radiography systems with regard to radiopacity of root canal filling materials. Int Endod J 2012. doi: 10.1111/j.1365-2591.2012.02026.x. [Epub ahead of print]. 20. Petersson A, Axelsson S, Davidson T, et al. Radiological diagnosis of periapical bone tissue lesions in endodontics: a systematic review. Int Endod J 2012. doi: 10.1111/j. 1365-2591.2012.02034.x. [Epub ahead of print]. 21. Goldman M, Pearson AH, Darzenta N. Endodontic success: who’s reading the radiograph? Oral Surg Oral Med Oral Pathol 1972;33:432–7. 22. Tewary S, Luzzo J, Hartwell G. Endodontic radiography: who is reading the digital radiograph? J Endod 2011;7:919–21. 23. Cleophas TJ, Cleophas TF. Artificial intelligence for diagnostic purposes: principles, procedures and limitations—clinical chemistry and laboratory medicine. CCLM/ FESCC 2010;48:159–65. 24. McMillen T, Simen P, Behseta S. Hebbian learning in linear-nonlinear networks with tuning curves leads to near-optimal, multi-alternative decision making. Neural Netw 2011;24:417–26. 25. Zou J, Han Y, So SS. Overview of artificial neural networks. Methods Mol Biol 2008; 458:15–23. 26. Mario MC, Abe JM, Ortega NR, Del Santo M Jr. Paraconsistent artificial neural network as auxiliary in cephalometric diagnosis. Artif Organs 2010;34: E215–21. 27. Bas B, Ozgonenel O, Ozden B, Bekcioglu B, Bulut E, Kurt M. Use of artificial neural network in differentiation of subgroups of temporomandibular internal derangements: a preliminary study. J Oral Maxillofac Surg 2012;70:51–9. 28. Saghiri MA, Asgar K, Boukani KK, et al. A new approach for locating the minor apical foramen using an artificial neural network. Int Endod J 2012;45:257–65. 29. Martos J, Ferrer-Luque CM, Gonzalez-Rodriguez MP, Castro LA. Topographical evaluation of the major apical foramen in permanent human teeth. Int Endod J 2009;42: 329–34. 30. Martos J, Lubian C, Silveira LF, Suita de Castro LA, Ferrer Luque CM. Morphologic analysis of the root apex in human teeth. J Endod 2010;36:664–7. 31. Pratten DH, McDonald NJ. Comparison of radiographic and electronic working lengths. J Endod 1996;22:173–6. 32. Green D. A stereo-binocular microscopic study of the root apices and surrounding areas of 100 mandibular molars: preliminary study. Oral Surg Oral Med Oral Pathol 1955;8:1298–304. 33. Nekoofar MH, Ghandi MM, Hayes SJ, Dummer PM. The fundamental operating principles of electronic root canal length measurement devices. Int Endod J 2006;39: 595–609. 34. ElAyouti A, Weiger R, Lost C. The ability of root ZX apex locator to reduce the frequency of overestimated radiographic working length. J Endod 2002;28: 116–9. 35. Rajan V, Varghese B, van Leeuwen TG, Steenbergen W. Review of methodological developments in laser Doppler flowmetry. Lasers Med Sci 2009;24:269–83. 36. Finkler T. The Effect of Insulating K-type Files on Accuracy and Reliability as Used in Two Electronic Apex Locators. Master’s thesis. Richmond, VA: Virginia Commonwealth University; 2011. 37. Nicotera P, Bellomo G, Orrenius S. Calcium-mediated mechanisms in chemically induced cell death. Annu Rev Pharmacol Toxicol 1992;32:449–70.
Artificial Neural Network in Endodontics
1133
Basic Research—Technology 38. Williams CB, Joyce AP, Roberts S. A comparison between in vivo radiographic working length determination and measurement after extraction. J Endod 2006; 32:624–7. 39. Dummer PM, McGinn JH, Rees DG. The position and topography of the apical canal constriction and apical foramen. Int Endod J 1984;17:192–8. 40. Olson AK, Goerig AC, Cavataio RE, Luciano J. The ability of the radiograph to determine the location of the apical foramen. Int Endod J 1991;24:28–35. 41. Forsberg J. Radiographic reproduction of endodontic "working length" comparing the paralleling and the bisecting-angle techniques. Oral Surg Oral Med Oral Pathol 1987;64:353–60. 42. Surmont P, D’Hauwers R, Martens L. Determination of tooth length in endodontics. Rev Belge Med Dent 1984;47:30–8.
1134
Saghiri et al.
43. Patel JL, Goyal RK. Applications of artificial neural networks in medical science. Curr Clin Pharmacol 2007;2:217–26. 44. Wei C, Wang C, Huang S. Using threshold method to separate the edge, coating and body of tongue in automatic tongue diagnosis. In: Networked Computing and Advanced Information Management (NCM), 2010 Sixth International Conference on. IEEE Conference Publications; 2010: 653-6. Available at http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=5573255&url=http%3A% 2F%2Fieeexplore.ieee.org%2Fstamp%2Fstamp.jsp%3Ftp%3D%26arnumber%3D557 3255. Accessed June 13, 2012 45. Okumura E, Kawashita I, Ishida T. Computerized analysis of pneumoconiosis in digital chest radiography: effect of artificial neural network trained with power spectra. J Digit Imaging 2011;24:1126–32.
JOE — Volume 38, Number 8, August 2012
The Reliability of Artificial Neural Network in Locating Minor Apical Foramen: A Cadaver Study Mohammad Ali Saghiri, BSc, MSc, PhD,* Franklin Garcia-Godoy, DDS, MS, PhD,† James L. Gutmann, DDS, PhD,‡ Mehrdad Lotfi, DMD, MSc,§ and Kamal Asgar, BSc, MSc, PhDjj Abstract Introduction: The purpose of this study was to evaluate the accuracy of the artificial neural network (ANN) in a human cadaver model in an attempt to simulate the clinical situation of working length determination. Methods: Fifty single-rooted teeth were selected from 19 male cadavers ranging in age from 49–73 years. Access cavities were prepared, a file was placed in the canals, and the working length was confirmed radiographically by endodontists. The location of the file in relation to the minor apical foramen was categorized as long, short, and exact by the ANN, by endodontists before extraction, and stereomicroscopically after extraction. The results were compared by using Friedman and Wilcoxon tests. The significance level was set at P <.05. Results: The Friedman test revealed a significant difference among groups (P < .001). There were significant differences between data obtained from endodontists and ANN (P = .001) and data obtained from endodontists and real measurements by stereomicroscope after extraction (P < .002). The correct assessment by the endodontists was accurate in 76% of the teeth. ANN determined the anatomic position correctly 96% of the time. The confidence interval for the correct result was 64.16–87.84 for endodontists and 90.57–101.43 for ANN. Conclusions: ANN was more accurate than endodontists’ determinations when compared with real working length measurements by using the stereomicroscope as a gold standard after tooth extraction. The artificial neural network is an accurate method for determining the working length. (J Endod 2012;38:1130–1134)
Key Words Apical foramen, artificial neural network, cadaver, root canal treatment, working length
C
orrect working length (WL) determination is an essential step in achieving successful root canal treatment outcomes. Instrumentation beyond the apical foramen (1), flare-ups (2), periapical foreign body reactions (3), and poor microbial control (4) are the result of poor WL determination and control. In 1930 Grove (5) stated that the proper point to which root canals should be cleaned and filled is the junction of the dentin and the cementum and that the pulp should be severed at the point of its union with the periodontal membrane. The cementodentinal junction (CDJ) is the anatomic and histologic landmark where the periodontal ligament begins and the pulp ends. Although the CDJ is a practical termination point for the preparation and obturation of the root canal, this anatomic point cannot be determined radiographically. However, for decades academic programs taught that the root canal procedures of shaping, enlarging, and obturation should be at or slightly short of the apical location. At the First World Endodontic Conference held in Philadelphia in 1953, 2 of the universal principles agreed upon were the following: (1) Traumatic injury to the surrounding (periapical) soft tissue should be avoided at all times. To this end, instrumentation stops should be used, and instruments should be confined entirely with the root canal. (2) The canal filling should seal the apical foramen, and if the apical millimeter or so of the canal is filled with healthy living tissue, the root canal filling should terminate at this level rather than at the apical foramen (6). An in vivo histologic study found that the most favorable histologic conditions were when the instrumentation and obturation remained short of the apical constriction and that extruded gutta-percha and sealer always caused a severe inflammatory reaction despite the absence of pain (7). The ability to achieve this position predictably is based on establishing a sound WL for each root. There are various methods for locating the apical foramen and determining the WL, including radiographic, digital tactile sense, and patients’ response to a file or paper point (8). Although little documentation exists to support the digital and patients’ response techniques (9), radiography is an accepted, commonly used method (10, 11). Some original radiographic methods requiring formulas, such as those of Bramante and Berbert (12), have been largely abandoned. Electronic apex locators (EALs) are very useful adjuncts during root canal procedures for locating the apical foramen (13), and recently, cone-beam computed tomography has also been determined to be an accurate means for determining the WL (14). The principles behind the first electronic device for root length determination were presented by Custer (15) in 1918 and commercialized by Sunada (16) in 1962 by using a device that registered consistent values in electrical resistance between an instrument in a root canal and an electrode on the oral mucous membrane (17). When reviewing the vast amount of literature that has addressed the efficacy and
From the *Department of Dental Material, Islamic Azad University (Tehran Branch), Tehran, Iran; †Bioscience Research Center, College of Dentistry, University of Tennessee Health Science Center, Memphis, Tennessee; ‡Department of Restorative Sciences, Baylor College of Dentistry, Texas A&M Health Science Center, Dallas, Texas; §Research Center for Pharmaceutical Nanotechnology and Department of Endodontics, Dental Faculty, Tabriz University (Medical Sciences), Tabriz, Iran; and k Department of Dental and Biological Materials, University of Michigan, Ann Arbor, Michigan. Address requests for reprints to Dr Mohammad Ali Saghiri, Assistant Professor, Department of Dental Material, Kamal Asgar Research Center (KARC) and Dental School, Azad University (Tehran Branch), P.O. Box 14665-1445, Tehran, Iran. E-mail address: [email protected] 0099-2399/$ - see front matter Copyright ª 2012 American Association of Endodontists. doi:10.1016/j.joen.2012.05.004
1130
Saghiri et al.
JOE — Volume 38, Number 8, August 2012
Basic Research—Technology accuracy of EALs, some in vitro experimental models give greater accuracy than can be achieved clinically (13, 18). No individual technique, however, is truly satisfactory in determining root canal WL (11). Digital radiography used for WL determination has the advantages of lower radiation exposure to the patient, ability to manipulate the digitized image for better observation, ease of radiographic data storage, and instant display of the image (19). The quality of the image is probably of utmost significance in endodontics because it facilitates accurate interpretation of root and canal morphology (20). However, clinical errors in radiographic interpretation by using standard films (21) and digital radiographs (22) create a need for an objective, computer-based analysis of WL. An artificial neural network (ANN) is a mathematical computerbased model inspired by the structure and/or functional aspects of biological neural networks in the brain. ANN is a decision-making system and helps the diagnostic procedure used for prediction of different elements from radiographs. ANNs are computer models with a massive parallel structure that imitate the human brain (23). They consist of nonlinear calculation elements, which represent neurons organized in several highly interconnected layers inspired by biological neural networks (24). The basic idea is that even though artificial neurons have no intelligence individually, when interconnected, they might be able to duplicate aspects of the human brain combined with the computational power of computers. ANNs have proved to be extremely helpful in solving problems with high computational complexity. ANNs have also been established as promising alternatives to statistical discriminate analysis because they can synthesize a considerable amount of information (or variables) without requiring statistical modeling of the problem (25). Because medical diagnosis is one of typical pattern recognition problem, previous applications have inspired the use of ANNs in the dentistry field such as cephalometric diagnosis (26) and differentiation of subgroups of temporomandibular internal derangements (27). Recently an investigation (28) showed that ANNs can act as a second opinion to locate the apical foreman on radiographs to enhance the accuracy of WL determination by radiography and can function as a decision-making system in various similar clinical situations. The lack of objectiveness in determination of WL coupled with high levels of variability among clinicians about the location of the file on the radiograph can impact greatly on the treatment rendered, patient discomfort, and successful outcomes. Therefore, the aim of this study was to evaluate the accuracy of an objective method of determining the WL in a human cadaver model and compare it with endodontists’ interpretations and microscopic realities.
Materials and Methods The research protocol was approved by the Research Ethics Committee of Kamal Asgar Research Center (protocol no. KARC/ 10B2011-70-20). Fifty single-rooted teeth with curvatures <30 in the apical region were selected from fresh cadavers that were males ranging in age between 49 and 73 years. This part of the study was similar to our previous study (28). Briefly, access cavities were prepared, and WLs were determined by placing a #20 K-file (Dentsply Maillefer, Ballaigues, Switzerland) in the root canal and exposing radiographs by using a Rinn XCP (Dentsply Rinn, Elgin, IL) and a photostimulable phosphorus system (Digora, Soredex, Tuusula, Finland). The files were fixed in the tooth by using cyanoacrylate glue (Pattex; Henkel AG & Co, Duesseldorf, Germany). The teeth were extracted, and after a gentle wash with normal saline the apical portions of the teeth were assessed by using a stereomicroscope (Nikon Corporation, JOE — Volume 38, Number 8, August 2012
Tokyo, Japan) equipped with a CCD camera at 20 to determine the position of the files. The positions were classified as follows: (1) File tip was not visible under the stereomicroscope (short). (2) The file tip was visible under the stereomicroscope, but it was not visible more than 0.3 mm beyond the minor foramen (exact). (3) The file was visible under the stereomicroscope and was more than 0.3 mm beyond the minor foramen (long). The prototype based on multilayer perceptron ANN model in combination with the Otsu method and wave let theory that had been used in previous study (28) was used to determine the WL (Fig. 1). Two endodontists evaluated the position of the files according to the position of the apical foramen on the radiographs, which visualized the root apices under radiographs by using a Rinn XCP and a photostimulable phosphorus system. If there were different opinions regarding the position of the file, another endodontist evaluated the radiographs. At least 2 similar opinions were considered as the final decision. The results of the endodontists’ evaluation, the ANN, and stereomicroscope assessment were compared by using Friedman and Wilcoxon tests. The significance level was set at P <.05.
Results The Friedman test revealed a significant difference among groups (P < .001). The Wilcoxon test was used to reveal any significant differences between specific groups. There were significant differences between data obtained from endodontists and ANN (P = .001) and data obtained from the endodontists and real measurements obtained from the stereomicroscopic assessment (P < .002) (Fig. 2). However, there were no significant differences between measurements of root canal length when ANN was compared with real measurements after extraction (P = .74). The determination of the minor anatomic constriction correctly by endodontists occurred 76% of the time, whereas the ANN identified this position correctly 96% of the time. Confidence interval for correct results was 64.16–87.84 for endodontists and 90.57–101.43 for ANN.
Discussion In most teeth requiring root canal procedures, the apical foramen is located on the lateral surface of the root at an average distance of 0.36 mm for anterior teeth and 0.48 mm for posterior teeth. Martos et al (29) found that the average distance between the apical foramen and the anatomic apex was 0.69 mm. This distance, however, was different in posterior teeth (0.82 mm) when compared with anterior teeth (0.39 mm); thus, in 60% of cases, the apical foramen is located on a lateral surface of the root and in 40% of cases on the anatomic apex. The same authors illustrated that the lateral position of the apical foramen has been detected in 43% of posterior teeth and in 17% of anterior teeth (30). WL can be determined with radiographs, tactile sensation, and EALs. However, radiographs are subject to distortion, magnification, interpretation variability, and lack of 3-dimensional representation. As a result, WL is generally measured about 0.5–1 mm short of the radiographic apex. Pratten and McDonald (31) showed that considering the position of the apical constriction located 1 mm short of the radiographic apex will result in an underestimation of WL. Vertical and horizontal cone angulations, film position, tooth inclination, and film processing issues could also influence WL determination by using radiographs (21). In some cases, the apical foramen might be located on the lateral root surface up to 3.5 mm from the radiographic apex (32). In such teeth, if the canal terminates coronal to the anatomic
Artificial Neural Network in Endodontics
1131
Basic Research—Technology
Figure 1. (A) Sample of radiograph (exact) and (B) stereomicroscope (long) and (C) snapshot of ANN software.
apex and in the plane of the film, the radiographic appearance will be short, and any adjustment will result in an overestimation of the WL. Numerous studies have reported on the accuracy of EALs for determining the location of the apical foramen. EALs can accurately determine the WL in 75.0%–96.5% of the root canals with mature apices (13, 33). In the presence of conductive restorations, severely undermined caries, serous, purulent, or hemorrhagic exudates, wide-canal or a wide-open apex WL determination is not accurate (34). In the present study EAL methods of WL determination were not used for 3 reasons: first, the blood pressure of cadavers was not the same as in vivo studies (35); second, the temperatures of corpses were lower than clinical scenario, which may impact on results (36); and third, in the presence of cellular death, the electric potential of cells is vastly different from that in viable cells (37), which may influence conductivity. The validity of measurements made with an in vitro experimental protocol, that is, the extent to which they depict the clinical accuracy of EALs, is unknown (17). However, they are able to reproduce the clinical condition of EAL use and facilitate an objective examination of several variables that are not practical in in vivo studies (33). WL determination by using radiographs is indispensable part of root canal treatment (38). The use of a combination of methods to determine an accurate WL may be more successful than relying on just 1 method (39). A parallel radiograph seems to be more accurate for determining the location of the apical foramen in human teeth during root canal
instrumentation (40) and a better reproduction of the distance between the apex and a contrasting subject (eg, a root canal instrument) (41). However, it has been stated that the accurate determination or even estimation of the apical canal constriction is not possible with radiography because of anatomic variations or errors in projection (34, 42). A neural network is a combination of soft and hardware simulation of a biological brain. The aim of a neural network is to learn to recognize patterns from entered data. Once the neural network has been trained on entered data, it can make predictions by detecting similar patterns in future data. Neural networks are a branch of biomedical engineering known as artificial intelligence. Other branches include genetic algorithms, expert systems, fuzzy logic, and chaos theory (43). ANN can be considered as a black box that is able to predict an output pattern when it recognizes a given input pattern. This network must first be "trained" by having it process a large number of input patterns and showing it what output resulted from each input pattern. Once trained, the network is able to recognize similarities when presented with a new input pattern, resulting in a predicted output pattern (27). In the current study we used a prototype that trained in a previous study (28) and the Otsu feature as a built-in. The Otsu method was used to automatically perform shape-based image thresholding histogram. This method can separate the teeth from surrounding tissues on the basis of differences in gray scales on radiographs (28). The teeth and surrounding tissues have many distinct features such as color and intensity. The algorithm assumes that the image to be thresholded
Figure 2. The plot of validity of neural network (A) and endodontists (B) showed that the correct identification for endodontists was 76%, whereas ANN determination was correct 96% of the time.
1132
Saghiri et al.
JOE — Volume 38, Number 8, August 2012
Basic Research—Technology contains 2 classes of pixels. The white color of the teeth produces 1 pinnacle in the histogram, and gray scales lower than the peaks (threshold) are removed by first determining a threshold independently. Only differences between the white color of the teeth and surrounding tissues remain. The Otsu method searches for a threshold that minimizes the intraclass variance, defined as a weighed sum of variances of the 2 classes (44). The major advantage of ANNs is their ability to be used as an arbitrary function approximation mechanism that learns from observed data. It should be mentioned that ANN systems are close dependent to pool of data that are used for training, so if we increase the pool of data for training system, logically we should be more accurate (23). ANN is capable of learning important relationships from a set of data and applying this knowledge to evaluate new cases. Also it modifies its behavior (trains) by adjusting the strength or weights of the connections until its own output converges to the known correct output. Once trained, the network can evaluate a new case of input values by applying the weights learned from the data set with which it was trained (45). The network can generalize from previous cases to evaluate cases not previously seen. It has been stated that ANN can act as a second opinion for locating the apical foramen in an in vitro situation (28). The present study showed that the cadavers can be used as a model to examine questions regarding the soft tissue interferences with the function and application of the ANN in clinical situations.
Conclusion Within the limitations of the present experiments, the following could be concluded. ANN is an accurate method for determining the WL in human cadavers. Incorporation of surrounding tissue does not have a deleterious effect on ANN decision-making. ANN determined the anatomic position correctly 96% of the time. Observations of ANN results suggest that it might be considered as an alternative for EAL or even confirmation of the results obtained with an EAL; however, further investigation is recommended.
Acknowledgments This project is dedicated to Sophie Asgar (late wife of Professor Asgar), who died during experimental part of present study. In addition, we thank Professor Mohammad Aeinehchi for providing all facilities for working on cadavers, Professors Shahram Azimi and Payman Mehrvazfar for making score and contributing to analysis, and Dr Houtan Aghili from IBM, New York for his invaluable help and re-check of our network and program. The authors deny any conflicts of interest related to this study.
References 1. Smith CS, Setchell DJ, Harty FJ. Factors influencing the success of conventional root canal therapy: a five-year retrospective study. Int Endod J 1993;26:321–33. 2. Georgopoulou M, Anastassiadis P, Sykaras S. Pain after chemomechanical preparation. Int Endod J 1986;19:309–14. 3. Nair MK, Nair UP. Digital and advanced imaging in endodontics: a review. J Endod 2007;33:1–6. 4. Chugal NM, Clive JM, Spangberg LS. Endodontic infection: some biologic and treatment factors associated with outcome. Oral Surg Oral Med Oral Path Oral Radio Endod 2003;96:81–90. 5. Grove C. Why canals should be filled to the dentinocemental junction. J Am Dent Assoc 1930;17:293–6. 6. Grossman LI. Transactions of the World Conference on Endodontics. Philadelphia: University of Pennsylvania; 1953.
JOE — Volume 38, Number 8, August 2012
7. Ricucci D, Langeland K. Apical limit of root canal instrumentation and obturation, part 2: a histological study. Int Endod J 1998;31:394–409. 8. Seidberg B, Alibrandi B, Fine H, Logue B. Clinical investigation of measuring working lengths of root canals with an electronic device and with digital-tactile sense. J Am Dent Assoc 1975;90:379–87. 9. Gutmann JL, Lovdahl PE. Problem Solving in Endodontics: Prevention, Identification, and Management, 5th ed. Maryland Heights, MO: Elsevier/Mosby; 2011. 10. Powell-Cullingford AW, Pitt Ford TR. The use of E-speed film for root canal length determination. Int Endod J 1993;26:268–72. 11. Gutmann JL, Leonard JE. Problem solving in endodontic working-length determination. Compend Contin Educ Dent 1995;16:288. 290, 293–284 passim; quiz 304. 12. Bramante CM, Berbert A. A critical evaluation of some methods of determining tooth length. Oral Surg Oral Med Oral Pathol 1974;37:463–73. 13. Gordon MP, Chandler NP. Electronic apex locators. Int Endod J 2004;37:425–37. 14. Janner SF, Jeger FB, Lussi A, Bornstein MM. Precision of endodontic working length measurements: a pilot investigation comparing cone-beam computed tomography scanning with standard measurement techniques. J Endod 2011; 37:1046–51. 15. Custer L. Exact methods of locating the apical foramen. J Natl Dent Assoc 1918;5: 815–9. 16. Sunada I. New method for measuring the length of the root canal. J Dent Res 1962; 41:375–87. 17. Kim E, Lee SJ. Electronic apex locator. Dent Clin North Am 2004;48:35–54. 18. Venturi M, Breschi L. A comparison between two electronic apex locators: an ex vivo investigation. Int Endod J 2007;40:362–73. 19. Akcay I, Ilhan B, Dundar N. Comparison of conventional and digital radiography systems with regard to radiopacity of root canal filling materials. Int Endod J 2012. doi: 10.1111/j.1365-2591.2012.02026.x. [Epub ahead of print]. 20. Petersson A, Axelsson S, Davidson T, et al. Radiological diagnosis of periapical bone tissue lesions in endodontics: a systematic review. Int Endod J 2012. doi: 10.1111/j. 1365-2591.2012.02034.x. [Epub ahead of print]. 21. Goldman M, Pearson AH, Darzenta N. Endodontic success: who’s reading the radiograph? Oral Surg Oral Med Oral Pathol 1972;33:432–7. 22. Tewary S, Luzzo J, Hartwell G. Endodontic radiography: who is reading the digital radiograph? J Endod 2011;7:919–21. 23. Cleophas TJ, Cleophas TF. Artificial intelligence for diagnostic purposes: principles, procedures and limitations—clinical chemistry and laboratory medicine. CCLM/ FESCC 2010;48:159–65. 24. McMillen T, Simen P, Behseta S. Hebbian learning in linear-nonlinear networks with tuning curves leads to near-optimal, multi-alternative decision making. Neural Netw 2011;24:417–26. 25. Zou J, Han Y, So SS. Overview of artificial neural networks. Methods Mol Biol 2008; 458:15–23. 26. Mario MC, Abe JM, Ortega NR, Del Santo M Jr. Paraconsistent artificial neural network as auxiliary in cephalometric diagnosis. Artif Organs 2010;34: E215–21. 27. Bas B, Ozgonenel O, Ozden B, Bekcioglu B, Bulut E, Kurt M. Use of artificial neural network in differentiation of subgroups of temporomandibular internal derangements: a preliminary study. J Oral Maxillofac Surg 2012;70:51–9. 28. Saghiri MA, Asgar K, Boukani KK, et al. A new approach for locating the minor apical foramen using an artificial neural network. Int Endod J 2012;45:257–65. 29. Martos J, Ferrer-Luque CM, Gonzalez-Rodriguez MP, Castro LA. Topographical evaluation of the major apical foramen in permanent human teeth. Int Endod J 2009;42: 329–34. 30. Martos J, Lubian C, Silveira LF, Suita de Castro LA, Ferrer Luque CM. Morphologic analysis of the root apex in human teeth. J Endod 2010;36:664–7. 31. Pratten DH, McDonald NJ. Comparison of radiographic and electronic working lengths. J Endod 1996;22:173–6. 32. Green D. A stereo-binocular microscopic study of the root apices and surrounding areas of 100 mandibular molars: preliminary study. Oral Surg Oral Med Oral Pathol 1955;8:1298–304. 33. Nekoofar MH, Ghandi MM, Hayes SJ, Dummer PM. The fundamental operating principles of electronic root canal length measurement devices. Int Endod J 2006;39: 595–609. 34. ElAyouti A, Weiger R, Lost C. The ability of root ZX apex locator to reduce the frequency of overestimated radiographic working length. J Endod 2002;28: 116–9. 35. Rajan V, Varghese B, van Leeuwen TG, Steenbergen W. Review of methodological developments in laser Doppler flowmetry. Lasers Med Sci 2009;24:269–83. 36. Finkler T. The Effect of Insulating K-type Files on Accuracy and Reliability as Used in Two Electronic Apex Locators. Master’s thesis. Richmond, VA: Virginia Commonwealth University; 2011. 37. Nicotera P, Bellomo G, Orrenius S. Calcium-mediated mechanisms in chemically induced cell death. Annu Rev Pharmacol Toxicol 1992;32:449–70.
Artificial Neural Network in Endodontics
1133
Basic Research—Technology 38. Williams CB, Joyce AP, Roberts S. A comparison between in vivo radiographic working length determination and measurement after extraction. J Endod 2006; 32:624–7. 39. Dummer PM, McGinn JH, Rees DG. The position and topography of the apical canal constriction and apical foramen. Int Endod J 1984;17:192–8. 40. Olson AK, Goerig AC, Cavataio RE, Luciano J. The ability of the radiograph to determine the location of the apical foramen. Int Endod J 1991;24:28–35. 41. Forsberg J. Radiographic reproduction of endodontic "working length" comparing the paralleling and the bisecting-angle techniques. Oral Surg Oral Med Oral Pathol 1987;64:353–60. 42. Surmont P, D’Hauwers R, Martens L. Determination of tooth length in endodontics. Rev Belge Med Dent 1984;47:30–8.
1134
Saghiri et al.
43. Patel JL, Goyal RK. Applications of artificial neural networks in medical science. Curr Clin Pharmacol 2007;2:217–26. 44. Wei C, Wang C, Huang S. Using threshold method to separate the edge, coating and body of tongue in automatic tongue diagnosis. In: Networked Computing and Advanced Information Management (NCM), 2010 Sixth International Conference on. IEEE Conference Publications; 2010: 653-6. Available at http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=5573255&url=http%3A% 2F%2Fieeexplore.ieee.org%2Fstamp%2Fstamp.jsp%3Ftp%3D%26arnumber%3D557 3255. Accessed June 13, 2012 45. Okumura E, Kawashita I, Ishida T. Computerized analysis of pneumoconiosis in digital chest radiography: effect of artificial neural network trained with power spectra. J Digit Imaging 2011;24:1126–32.
JOE — Volume 38, Number 8, August 2012