Abstract
Objectives
Artificial intelligence (AI) techniques like convolutional neural network (CNN) are a promising breakthrough that can help clinicians analyze medical imaging, diagnose taurodontism, and make therapeutic decisions. The purpose of the study is to develop and evaluate the function of CNN-based AI model to diagnose teeth with taurodontism in panoramic radiography.
Methods
434 anonymized, mixed-sized panoramic radiography images over the age of 13 years were used to develop automatic taurodont tooth segmentation models using a Pytorch implemented U-Net model. Datasets were split into train, validation, and test groups of both normal and masked images. The data augmentation method was applied to images of trainings and validation groups with vertical flip images, horizontal flip images, and both flip images. The Confusion Matrix was used to determine the model performance.
Results
Among the 43 test group images with 126 labels, there were 109 true positives, 29 false positives, and 17 false negatives. The sensitivity, precision, and F1-score values of taurodont tooth segmentation were 0.8650, 0.7898, and 0.8257, respectively.
Conclusions
CNN’s ability to identify taurodontism produced almost identical results to the labeled training data, and the CNN system achieved close to the expert level results in its ability to detect the taurodontism of teeth.
Similar content being viewed by others
Explore related subjects
Discover the latest articles, news and stories from top researchers in related subjects.Avoid common mistakes on your manuscript.
Introduction
Taurodontism is a condition in which the pulp chamber of a tooth is substantially expanded and furcation displaces to the apical area. This developmental condition can be present in both healthy people and people with syndromes, including Down syndrome, ectodermal dysplasia, and Klinefelter syndrome, among others. Taurodontism can also be seen with anomalies such as Amelogenesis imperfecta, cleft palate, microdontia, and dens invaginatus [1, 2]. Taurodontism is thought to be caused by invagination failure in the Hertwig epithelial root sheath; however, its etiological causes are unknown [3].
Taurodontism can cause some difficulties during endodontic, orthodontic, and/or prosthetic procedures, and has been reported to increase pulp exposure during dental procedures [4]. It may cause difficulties in pulpectomy, canal preparation, and filling in endodontics, as the shifting of furcation to the apical triad may complicate the extraction of the effected tooth [5]. Taurodontism can appear unilaterally or bilaterally, and it might affect a single tooth or several teeth. Taurodontism most frequently affects molar teeth [1, 5,6,7] and may require diverse treatment approaches [6]. Teeth appear clinically normal in taurodontism [8], but it is usually detected by conventional radiographs (panoramic, intraoral, etc.) or cone-beam computed tomography (CBCT) [9].
The term “artificial intelligence” (AI) originated in the 1950s and refers to the concept of building machines capable of performing tasks that are typically performed by humans [10]. The field of Artificial Intelligence (AI) is concerned with teaching machines how to learn, problem-solve, and analyze in ways similar to human beings. Machine learning is a process for achieving AI by building algorithms to teach machine behavior through examples and emulation. Deep learning (DL) is an AI method used to program automated decision-making for various clinical tasks. DL techniques, created through the development of artificial neural networks, enable the automatic categorization of datasets and enables learning from data via multi-layer convolutional neural networks (CNN). By functioning like interconnected neurons in the human brain, CNN can learn adaptive image features and perform image classification [11].
AI techniques like CNN are a promising breakthrough because they can help clinicians analyze medical imaging, diagnose diseases, and make therapeutic decisions. AI studies using CNNs [12,13,14,15,16,17,18,19,20,21,22,23,24,25,26] have also made progress in the field of dentistry. The present study aims to develop and evaluate the function of CNN designed to diagnose teeth with taurodontism in panoramic radiography.
Materials and methods
Study design
In this study, automatic taurodont tooth segmentation models (CranioCatch, Eskisehir, Turkey) were produced in panoramic radiographs using a Pytorch implemented U-Net model. Eskisehir Osmangazi University Non-interventional Clinical Research Ethics Board (decision date, meeting number and decision number: 06.08.2019/14) authorized the study protocol. The principles of the Helsinki Declaration were followed in the study.
Data sources
Included in the study were 434 anonymous panoramic radiographs of patients over the age of 13 years obtained from January 2018 to January 2020 from the radiology archive of the Department of Oral and Maxillofacial Radiology of Inonu University Faculty of Dentistry. Gender differences were not considered. Radiographs were achieved using the Planmeca Promax 2D Panoramic System (Planmeca, Helsinki, Finland) using the 68 kVp, 14 mA, and 12 s. image acquisition parameters.
Labeled training data
A dentomaxillofacial radiologist, a pediatric dentist, and an endodontist, all three with at least 5 years of experience, identified the labeled training data for the panoramic radiographs using CranioCatch Annotation Tool (Eskisehir, Turkey). A tooth with an apically displaced pulp chamber which did not show the usual constriction of the pulp at the cementoenamel junction and had an apically displaced furcation area was considered as a taurodont [27] (Fig. 1). The index value for each tooth identified in the taurodont. The ratio of the height of the pulp chamber to the length from the roof of the pulp chamber to the root apex multiplied by one hundred was recorded [28]. Molar teeth with a taurodontism index (TI) above 20 were included in the study (Table 1, Fig. 2). Experts were asked to draw polygonal boxes around all the taurodont teeth included in the study on panoramic radiographs independently. Standard monitor settings and lighting conditions were used for images. If the experienced dentists did not agree on a diagnosis, even after discussion, the diagnosis was considered inconsistent and the image was excluded.
Segmentation model
434 anonymized, mixed-sized panoramic radiography images were resized to 1024 × 512 diameter and cropped so that lower and upper jaw teeth appeared in the frame. Surrounding hazy images were cleared by applying contrast limited adaptive histogram equalization (CLAHE). Masked images were generated from labels, and datasets were split into train, validation, and test groups of both normal and masked images.
Training group 348 (1080 labels).
Training-mask group 348 (1080 labels).
Validation group 43 (126 labels).
Validation-mask group 43 (126 labels).
Test group 43 (126 labels).
Test-mask group 43 (126 labels).
The data augmentation method was applied to images of trainings and validation groups with vertical flip images, horizontal flip images, and both flip images. After the data augmentation, the images of groups were as follows:
Training group 348 (1080 labels) × 4 = 1392 (4320 labels).
Training-mask group 348 (1080 labels) × 4 = 1392 (4320 labels).
Validation group 43 (126 labels) × 4 = 172 (504 labels).
Validation-mask group 43 (126 labels) = 172 (504 labels).
Test group 43 (126 labels).
Test-mask group 43 (126 labels).
Deep-convolutional neural network (CNN) architecture
U-Net is a fully convolutional neural network whose main purpose is to segment medical images by estimating the class of each pixel in the image. As its name suggests, U-net architecture is “U” shaped and divided into three parts: the encoding or down-sampling path, the bottleneck, and the decoding and up-sampling path. The coding path consists of many coding blocks. After each block, the map feature numbers multiply so that the model can efficiently learn complex structures. The bottleneck between the down-sampling and up-sampling paths consists of two convolutional layers. The essence of the U-net architecture lies in the decoding part. The U-Net model works incrementally by encoding the input and finding its deep features, which are then decoded using deconvolution and combined with features of the same dimensions into a channel [29] (Fig. 3).
Training
Python, an open-source programming language (version 3.6.1; Python Software Foundation, Wilmington, DE, USA), was used to develop the AI algorithm. The PyTorch library was used for creating the AI algorithm within the U-Net network. The training method was applied using computer equipment from Eskisehir Osmangazi Faculty of Dentistry Dental-AI Laboratory, including a Dell PowerEdge T640 Calculation Server (Dell Inc., Texas, USA), a Dell PowerEdge T640 GPU Calculation Server (Dell Inc., Texas, USA), and a Dell PowerEdge R540 Storage Server (Dell Inc., Texas, USA). The taurodont tooth segmentation model with Pytorch U-Net was trained with 600 epochs and the best model 548th epoch was saved. An AI model (CranioCatch, Eskisehir, Turkey) was used to automatically segment the taurodont teeth on panoramic radiographs (Fig. 4).
Assessment of model performance
The Confusion Matrix was used to determine the model performance. The metrics used to assess the performance of the taurodont tooth segmentation model were as follows:
True Positive (TP): The number of accurately segmented taurodont teeth.
False Positive (FP): The number of segmented teeth with no taurodontism.
False Negative (FN): The number of teeth with taurodont not segmented.
The performance metrics of the model were determined according to the formulas as follows:
Results
A total of 126 labeled teeth with taurodontism (hypotaurodont; n = 39, mesotaurodont; n = 30, hypertaurodont; n = 57) in 43 panoramic radiographs were determined as the test group (Table 2). The mean TI scores of hypotaurodont, mesotaurodont, and hypertaurodont teeth in the test group were found to be 26.30 ± 2.33, 34.61 ± 3.25, and 82.55 ± 24.34, respectively (Table 3).
Among the 43 test group images with 126 labels, there were 109 TPs, 29 FPs, and 17 FNs (Table 4). The sensitivity, precision, and F1-score values of taurodont tooth segmentation were 0.8650, 0.7898, and 0.8257, respectively (Table 5 and Fig. 5).
Discussion
CNN systems have recently gained popularity in the medical field, particularly for the analysis of radiological images [30]. Many studies [13,14,15,16,17,18,19,20,21] have shown CNN to be successful at scanning radiographic images (periapical, bitewing, panoramic, CBCT, etc.) used in dentistry. Research conducted with conventional radiographic techniques [13, 15, 17, 18] using CNN to detect and number teeth described CNN as a promising technology for the future. Moderate and high success was achieved in studies evaluating panoramic radiographs of alveolar bone loss [12], teeth with periodontal problems [16], dental caries [21], apical lesions [22], vertical root fracture [23], root morphology [24, 25] and impacted teeth [26]. However, there was no study using CNN to evaluate teeth with taurodontism, so our study is a first in this respect. In this study, the CNN’s ability to identify taurodontism produced almost identical results to the labeled training data, and the CNN system achieved close to the expert level results in its ability to detect the taurodontism of teeth.
Artificial intelligence research in pediatric dentistry is a newly emerging field. In the study of You et al., clinically acceptable performance was obtained using the CNN model compared to a pediatric dentist experienced in detecting dental plaque in primary teeth [31]. Kılıç et al. concluded that fast regional-CNN for detecting and numbering the primary teeth of children in panoramic radiography images was a promising tool for automated charting of panoramic radiographs from children [13]. Calıskan et al. reported that a moderately deep CNN trained on a limited amount of image data showed satisfactory discriminatory ability to detect submerged molars on panoramic radiographs [26]. Mine et al. reported that CNN-based deep learning method is a promising approach to detect the presence of supernumerary teeth in early mixed dentition [32]. These studies can lead to the development and dissemination of pediatric dentistry-specific artificial intelligence by enabling collaboration between pediatric dentists and data scientists with the necessary expertise to apply big data sources to artificial intelligence.
This study used a CNN-based deep learning technique to diagnose taurodont teeth by analyzing panoramic images. Despite drawbacks, such as two-dimensional imaging, distortion, and superposition of anatomical structures, panoramic radiographs are used to diagnose dental anomalies since they allow the entire dentition to be seen on a single film. Moreover, particularly for pediatric dentists, radiographing is a tolerable procedure to the patients, time-efficient, and offers low radiation exposure in the diagnosis and treatment planning of anomalies, such as taurodontism, where more than one or symmetrical teeth may be affected [13]. However, difficulties may arise in interpreting panoramic images due to subjective evaluations and differences between observers [12, 13]. AI can be used as a decision-support system for physicians and help reduce diagnostic errors caused by fatigue, especially in institutions with high workloads, and may capture details that are overlooked by inexperienced physicians [12]. In the field of dentistry, AI can also provide objective data for image standardization, data archiving, as well as for early diagnosing and treatment planning [12].
Blumberg et al. [27] noted that taurodontism can be described metrically and that a precise mathematical tool can be invaluable in accomplishing this task. In this study, the ratio-based “taurodont index” (TI) value developed by Shifman and Chanannel for radiological classification with objective measurements was used for the diagnosis of taurodontism. Since one of the index variables is a measurement that includes the apex of the longest root, the lower limit of our study was determined as 13 age r to maintain the root/crown ratio of the teeth, complete the root length and ensure standardization.
Taurodontism affects a wide range of dental procedures. Teeth with taurodontism are limited due to attrition, especially in maturity, and the completion of root formation, which affects which teeth can be used, especially in younger patients who are more likely to be under the care of an orthodontist or pediatric dentist. In orthodontic and prosthetic considerations, taurodontism limits the root surface area available for anchoring, as the absence of a cervical narrowing would deprive the tooth of its supportive action against crown overload [33]. Furthermore, taurodontism complicates endodontic therapy since it can cause problems during instrumentation and obturation in root canal treatment. Endodontic treatment of a taurodont tooth becomes more complex as the furcation shifts towards the apical third [34]. A modified filling technique has been developed, which involves combining lateral compaction in the apical region with vertical compaction in the elongated pulp chamber [35]. Physicians must be aware of the complex canal system to manage it successfully. This study aimed to diagnose only teeth with taurodontism, because the CNN technique is based on incremental improvement phases, and the results of preliminary steps have employ for transfer learning. However, there are variations of teeth with taurodontism (hypotaurodont, mesotaurodont, and hypertaurodont) [3, 4], and several indices are used to classify them [4]. The taurodontic teeth indicated by the experts’ TI measurements were labeled and introduced to the artificial intelligence system in this study. The second stage involves using a deep learning system to score TIs and categorizing them based on the results. Without taking any measurements, physicians can get information on the vertical length of the pulp chamber and the predicted locations of the canal entrances. Furthermore, employing AI to determine the type of taurodontism can aid in the assessment of root canal treatment risks.
The segmentation of multi-rooted teeth is more complicated than single-rooted teeth. The maxillary molars teeth are multi-rooted and multiple roots can overlap with the maxillary sinus, making it difficult to automate the segmentation of these teeth [14]. Given that studies show that molar teeth [1, 5,6,7] and the maxilla [1, 7, 36] are the most common areas affected by taurodontism, the difficulty in segmenting molar teeth may also occur in the diagnosis of taurodontism. Since this study was the first study in which taurodont teeth were diagnosed, it was not evaluated whether there was a difference in the detection of taurodont teeth in the maxilla and mandible with AI, and in future studies, subgroup analysis for the maxilla and mandible could be performed.
Conclusion
In this study, CNN models performed well in the detection of taurodont teeth on panoramic radiographs. Although further improvements are needed for clinical applications, the CNN-based deep learning approach is a promising approach for supporting dentists in dental practice and plays a precious role in saving time. In future studies CNN could be extended to evaluate different types of taurodontism or to diagnose using different types of X-ray images, such as periapical, CBCT.
References
Weckwerth GM, Santos CF, Brozoski DT, Pagin CBS, O, Lauris JRP, et al. Taurodontism, root dilaceration, and tooth transposition: a radiographic study of a population with nonsyndromic cleft lip and/or palate. Cleft Palate Cranio-fac J. 2016;53:404–12.
Topal BG, Tiras M. Developmental anomalies affecting tooth roots. Dent Med J-R. 2020;2:111–26.
Manjunatha B, Kovvuru SK. Taurodontism: a review on its etiology, prevalence and clinical considerations. J Clin Exp Dent. 2010;2:1–5.
Jafarzadeh H, Azarpazhooh A, Mayhall JT. Taurodontism: a review of the condition and endodontic treatment challenges. Int Endod J. 2008;41:375–88.
Bronoosh P, Haghnegahdar A, Dehbozorgi M. Prevalence of taurodontism in premolars and molars in the South of Iran. J Dent Res Dent Clin Dent Prospects. 2012;6:21–4.
Umar E, Altun O, Dedeoglu N. The retrospectıve evaluation of taurodontism prevalence in patiens admitting Inonu Univercity Faculty of Dentistry. Cumhur Dent J. 2014;17:235–43.
Nalcacı R, Gorgun S, Karakaya M. Incidence of taurodontism in Turkish population. Turkiye Klinikleri J Dent Sci. 2000;6:178–82.
White SC, Pharoah MJ. Oral radiology-e-book: principles and interpretation. 5th ed. St. Louis, USA: Mosby; 2004.
Luan X, Ito Y, Diekwisch TGH. Evolution and development of Hertwig’s epithelial root sheath. Dev Dyn. 2006;235:1167–80.
Schwendicke FA, Samek W, Krois J. Artificial intelligence in dentistry: chances and challenges. J Dent Res. 2020;99:769–74.
Kim DH, MacKinnon T. Artificial intelligence in fracture detection: transfer learning from deep convolutional neural networks. Clin Radiol. 2018;73:439–45.
Kurt Bayrakdar S, Celik O, Bayrakdar IS, Orhan K, Bilgir E, Odabas A, et al. Success of artificial intelligence system in determining alveolar bone loss from dental panoramic radiography images. Cumhur Dent J. 2020;23:318–24.
Kılıc Coruh M, Bayrakdar IS, Celik O, Bilgir E, Orhan K, Aydın OB, et al. Artificial intelligence system for automatic deciduous tooth detection and numbering in panoramic radiographs. Dentomaxillofac Radiol. 2021;50:20200172.
Lee JH, Han SS, Kim YH, Lee C, Kim I. Application of a fully deep convolutional neural network to the automation of tooth segmentation on panoramic radiographs. Oral Surg Oral Med Oral Pathol Oral Radiol. 2020;129:635–42.
Tuzoff DV, Tuzova LN, Bornstein MM, Krasnov AS, Kharchenko NSI, et al. Tooth detection and numbering in panoramic radiographs using convolutional neural networks. Dentomaxillofac Radiol. 2019;48:20180051.
Thanathornwong B, Suebnukarn S. Automatic detection of periodontal compromised teeth in digital panoramic radiographs using faster regional convolutional neural networks. Imaging Sci Dent. 2020;50:169.
Chen H, Zhang K, Lyu P, Li H, Zhang L, Wu J, et al. A deep learning approach to automatic teeth detection and numbering based on object detection in dental periapical films. Sci Rep. 2019;9:1–11.
Yasa Y, Celik O, Bayrakdar IS, Pekince A, Orhan K, Akarsu S, et al. An artificial intelligence proposal to automatic teeth detection and numbering in dental bite-wing radiographs. Acta Odontol Scand. 2021;79:275–81.
Kunz F, Stellzig-Eisenhauer A, Zeman F, Boldt J. Evaluation of a fully automated cephalometric analysis using a customized convolutional neural network. J Orofac Orthop. 2020;81:52–68.
Orhan K, Bayrakdar IS, Ezhov M, Kravtsov A, Ozyurek T. Evaluation of artificial intelligence for detecting periapical pathosis on cone-beam computed tomography scans. Int Endod J. 2020;53:680–9.
Lee JH, Kim DH, Jeong SN, Choi SH. Detection and diagnosis of dental caries using a deep learning-based convolutional neural network algorithm. J Dent. 2018;77:106–11.
Ekert T, Krois J, Meinhold L, Elhennavy K, Emare R, Golla T, et al. Deep learning for the radiographic detection of apical lesions. J Endod. 2019;45:917–22.
Fukuda M, Inamoto K, Shibata N, Ariji Y, Yanashita Y, Kutsuna S, et al. Evaluation of an artificial intelligence system for detecting vertical root fracture on panoramic radiography. Oral Radiol. 2020;36:337–43.
Hiraiwa T, Ariji Y, Fukuda M, Kise Y, Nakata K, Katsumata A, et al. A deep-learning artificial intelligence system for assessment of root morphology of the mandibular first molar on panoramic radiography. Dentomaxillofac Radiol. 2019;48:20180218.
Jeon SJ, Yun JP, Yeom HG, Shin WS, Lee JH, Jeong SH, et al. Deep-learning for predicting C-shaped canals in mandibular second molars on panoramic radiographs. Dentomaxillofac Radiol. 2021;50:20200513.
Caliskan S, Tuloglu N, Celik O, Ozdemir C, Kizilaslan S, Bayrak S. A pilot study of a deep learning approach to submerged primary tooth classification and detection. Int J Comput Dent. 2021;24:1–9.
Blumberg JE, Hylander WL, Goepp RA. Taurodontism: a biometric study. Am J Phys Anthropol. 1971;34:243–55.
Shifman A, Chanannel I. Prevalence of taurodontism found in radiographic dental examination of 1,200 young adult Israeli patients. Community Dent Oral Epidemiol. 1978;6:200–3.
Olaf R, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. Springer, Cham, 2015.
Yasaka K, Akai H, Kunimatsu A, Kiryu S, Abe O. Deep learning with convolutional neural network in radiology. Jpn J Radiol. 2018;36:257–72.
You W, Hao A, Li S, Wang Y, Xia B. Deep learning-based dental plaque detection on primary teeth: a comparison with clinical assessments. BMC Oral Health. 2020;20:141.
Mine Y, Iwamoto Y, Okazaki S, Nakamura K, Takeda S, Peng TY, et al. Detecting the presence of supernumerary teeth during the early mixed dentition stage using deep learning algorithms: a pilot study. Int J Paediatr Dent. 2021. https://doi.org/10.1111/ipd.12946.
MacDonald D. Taurodontism. Oral Radiol. 2020;36:129–32.
Manjunatha BS, Kovvuru SK. Taurodontism: a review on its etiology, prevalence and clinical considerations. J Clin Exp Dent. 2010;2:e187–90.
Tsesis I, Shifman A, Kaufman AY. Taurodontism: an endodontic challenge: report of a case. J Endod. 2003;29:353–5.
Cakici F, Benkli Y, Cakıcı E. The prevalence of taurodontism in a north Anatolian dental patient subpopulation. Middle Black Sea J Health Sci. 2015;1:7–10.
Funding
This work has been supported by Eskişehir University Scientific Research Projects Coordination Unit under Grant number 202045E06.
Author information
Authors and Affiliations
Contributions
SD, ISB, and EFY conceived the ideas; GE, SD, and EFY collected the data; OC, EB, and ISB analyzed the data; SD, ISB, and GE: writing—original draft; ALFC, RJ, and KO: supervision.
Corresponding author
Ethics declarations
Conflict of interest
The authors declare that they have no conflict of interest.
Ethical approval
All the procedures followed were in accordance with the ethical standards of the responsible committee on human experimentation (institutional and national) and with the Helsinki Declaration of 1975, as revised in 2008 (5). Informed consent was obtained from all the patients for being included in the study. This article does not contain any studies with human or animal subjects performed by the any of the author.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
About this article
Cite this article
Duman, S., Yılmaz, E.F., Eşer, G. et al. Detecting the presence of taurodont teeth on panoramic radiographs using a deep learning-based convolutional neural network algorithm. Oral Radiol 39, 207–214 (2023). https://doi.org/10.1007/s11282-022-00622-1
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11282-022-00622-1