Artificial Intelligence in Positioning Between Tooth and Nerve on Panoramic Radiography

Artificial Intelligence in Positioning Between Tooth and Nerve on Panoramic Radiography | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Artificial Intelligence in Positioning Between Tooth and Nerve on Panoramic Radiography Eunhye Choi, Soohong Lee, Eunjae Jeong, Seokwon Shin, Sekyoung Youm, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-888103/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 8 You are reading this latest preprint version Abstract Determining the exact positional relationship between mandibular third molar (M3) and inferior alveolar nerve (IAN) is important for surgical extractions. Panoramic radiography is the most common dental imaging test. The purposes of this study were to develop an artificial intelligence (AI) model to determine two positional relationships (true contact and bucco-lingual position) between M3 and IAN when they were overlapped in panoramic radiographs and compare its performance with that of oral and maxillofacial surgery (OMFS) specialists. A total of 571 panoramic images of M3 from 394 patients was used for this study. Among the images, 202 were classified as true contact, 246 as intimate, 61 as IAN buccal position, and 62 as IAN lingual position. A deep convolutional neural network model with ResNet-50 architecture was trained for each task. We randomly split the dataset into 75% for training and validation and 25% for testing. Model performance was superior in bucco-lingual position determination (accuracy 0.76, precision 0.83, recall 0.67, and F1 score 0.73) to true contact position determination (accuracy 0.63, precision 0.62, recall 0.63, and F1 score 0.61). AI exhibited much higher accuracy in both position determinations compared to OMFS specialists. In determining true contact position, OMFS specialists demonstrated an accuracy of 52.68% to 69.64%, while the AI showed an accuracy of 72.32%. In determining bucco-lingual position, OMFS specialists showed an accuracy of 32.26% to 48.39%, and the AI showed an accuracy of 80.65%. Moreover, Cohen’s kappa exhibited a substantial level of agreement for the AI (0.61) and poor agreements for OMFS specialists in bucco-lingual position determination. Determining the position relationship between M3 and IAN is possible using AI, especially in bucco-lingual positioning. The model could be used to support clinicians in the decision-making process for M3 treatment. Dentistry Artificial intelligence deep learning inferior alveolar nerve mandibular third molar panoramic radiography Figures Figure 1 Figure 2 Introduction Mandibular third molar (M3) extraction is one of the most frequently performed surgical procedures in oral and maxillofacial surgery (OMFS). Among the complications following surgery, damage to the inferior alveolar nerve (IAN) is one of the most distressing, causing temporary or permanent neurosensory impairments in the lower lip and chin area at an incidence of 0.4–13.4% 1,2 . To avoid IAN damage, preoperative assessment of the position of the IAN in relation to the tooth is necessary. Panoramic radiography is used commonly to assess the relationship between M3 and IAN. Certain radiographic features such as darkening of the root and narrowing of the mandibular canal have been reported as risk factors for IAN injuries, although its clinical correlation was low 3 . Due to the development of cone-beam computerized tomography (CBCT), determination of positioning between the IAN and teeth has become more accurate, and CBCT is recommended before M3 extraction when the two aforementioned structures are superimposed on panoramic radiography 4 . However, the disadvantages of CBCT include higher radiation doses compared to two-dimensional imaging and the presence of image artifacts mainly produced by metal restorations 5 . Therefore, accurate methods diagnosing the relationship between M3 and IAN on panoramic radiography are necessary. After the diagnostic methods determine whether both structures are truly in contact or intimate, assessment whether M3 is positioned lingually or buccally to the IAN is necessary to determine the direction of insertion of the surgical instruments. Artificial intelligence (AI) models have reported excellent performance, mimicking the precision and accuracy of trained specialists in dentistry 6 . Various studies have applied AI algorithms to read panoramic radiographs for clinical conditions such as age estimation 7 , osteoporosis 8 , 9 , vertical root fracture 10 , automatic teeth detection and numbering 11 , apical lesions 12 , maxillary sinusitis 13 , detecting and segmenting the approximation of the inferior alveolar nerve and mandibular third molar 14 , periodontal bone loss 15 , gender determination 16 , and temporomandibular joint osteoarthritis 17 , 18 . However, there is an AI study on the positional relationship between M3 and IAN, focusing on contact or non-contact 4 . This study aimed to investigate the clinical use of an AI model developed to determinate the positional relationship between M3 and IAN from panoramic radiography using deep learning that compared the AI readings with those of OMFS specialists. Methods Materials. The written documentation of informed consent was waived and approved by the decision of the Institutional Review Board of Seoul National University Dental Hospital (ERI21004) and ethics committee approval for the study in the same institute was also obtained. All methods were performed in accordance with the relevant guidelines and regulation. Subjects were included retrospectively from an image database of patients who visited the Department of Oral and Maxillofacial Surgery at Seoul National University Gwanak Dental Hospital between January 2019 and December 2020. Patients who underwent both panoramic radiography (Kodak 8000 Digital Panoramic System, Trophy Radiologies, Carestream Health Inc., NY, USA) and CBCT (CS 9300, Carestream Health Inc., NY, USA) for M3 extraction with superimposition of M3 and IAN on the panoramic radiographs were selected. The patients consisted of 200 males and 194 females, with an age range of 20 to 72 years (mean ± SD age, 31.5 4 ± 9.96 y; range, 20 to 72 y). The panoramic images of 571 M3s from these patients were used in this study. AI model developments. The AI model was developed to evaluate two positional relationships between M3 and IAN. Experiment 1: Determination of the true contact position between M3 and IAN Experiment 2: Determination of the bucco-lingual position between M3 and IAN Panoramic images that appeared overlapped were classified as true contact and intimate according to the presence or absence of the cortical line of the IAN canal on CBCT (Fig. 1 A, B). Independently, the bucco-lingual positional relationship was also confirmed by CBCT (Fig. 1 C, D). Determination of the positional relationship based on the CBCT was performed by an OMFR specialist. Among the 571 images, 202 were classified as true contact, 246 as false contact, 61 as IAN buccal position, and 62 as IAN lingual position. Regions of Interest (ROI) were extracted from the panoramic radiograph manually in JPG format with a matrix size of 400 x 400 pixels. ResNet-50, mainly used for medical image classification 19 , is a substantially deeper and easier model to train compared to simple models such as VGGnet, and the core structure is a residual block 20 . Residual learning does not allow for error accumulation on the convolution layers but enables a better representation of the content in the convolution layers. By adopting a shortcut structure, the vanishing gradient issue is resolved 21 . Every image was resized to 224 x 224 pixels, and we randomly split the dataset into 75% for training and validation and 25% for testing. The model performance varies depending on the difficulty of data classification, so we performed 5 repeated experiments through random sampling. As a technical and strategic method to avoid overfitting, data augmentation was performed by image rotation ± 30 degrees, horizontal flipping, and brightness 20–80% for every mini-batch in training to compensate for the small number of data points to increase model robustness. In Experiment 1, a model was trained for 60 epochs with augmented data. The learning rate of the model was set to 1.0 × 10–4 and an Adam optimizer was used. In Experiment 2, training was progressed in 30 epochs with augmented data. In addition, the learning rate of the model was 1.0 × 10–4, and an Adam optimizer was used. Model and statistical analysis. Accuracy, precision, recall, F1 score, and AUC were calculated to evaluate each model performance. Accuracy is defined as the ratio of correct predictions. Precision is the ratio of true positives to true positives and false positives. Recall is the ratio of true positives to true positives and false negatives. F1 score is a harmonic mean of precision and recall: (2 × precision × recall)/(precision + recall), and AUC is the area under the ROC curve. The confidence intervals of AUC were found by bootstrapping with 1,000 test sets sampled with replacement. For evaluation of AI clinical usability, the results between OPG reads by the AI and six OMFS specialists were compared. Accuracy, sensitivity, and specificity were calculated for diagnostic performance, and Cohen’s kappa was calculated to estimate the strength of agreement. Among the random samplings, the same test dataset with the highest accuracy was selected for comparison with specialists in each experiment. Python programming language (v. 3.8.5), Tensorflow (v. 2.5.0), and a graphics card (GeForce RTX 3090) were used for analysis. Data availability The data that support the findings this study are available from the authors upon reasonable request. Results Table 1 shows the model performance in each experiment. The average accuracy, precision, recall, and F1 score were 0.63, 0.62, 0.63, and 0.61, respectively, in Experiment 1, with true contact position determination between M3 and IAN. The average accuracy, precision, recall, and F1 score were 0.76, 0.83, 0.67, and 0.73, respectively, in Experiment 2, with bucco-lingual position determination between M3 and IAN. The overall model performance was superior in Experiment 2 compared to Experiment 1. Table 1 Model performance of five random samplings in each experiment Experiment 1, true contact position between M3 and IAN Work Accuracy Precision Recall F1 score AUC 1 0.72 0.72 0.55 0.63 0.75 2 0.55 0.54 0.74 0.62 0.59 3 0.67 0.65 0.59 0.62 0.67 4 0.61 0.55 0.68 0.61 0.66 5 0.60 0.62 0.57 0.60 0.64 Average 0.63 0.62 0.63 0.61 0.66 Experiment 2, bucco-lingual position between M3 and IAN 1 0.77 0.82 0.78 0.80 0.88 2 0.81 0.86 0.75 0.80 0.91 3 0.74 0.67 0.77 0.71 0.75 4 0.77 0.89 0.57 0.70 0.79 5 0.68 0.89 0.47 0.62 0.80 Average 0.76 0.83 0.67 0.73 0.83 M3, mandibular third molar; IAN, inferior alveolar nerve; AUC, Area under the ROC curve. The comparison of sensitivities and specificities between AI and OMFS specialists in each experiment is shown in Fig. 2 . AI exhibited 72.32% accuracy in Experiment 1 and 80.65% in Experiment 2, but the highest accuracy among OMFR specialists in each experiment was 69.64% and 51.61%, respectively (Table 2 ). Cohen’s kappa of AI was highest in Experiment 2 and showed a substantial level of agreement (0.61), but those of OMFS specialists exhibited a slight to fair level of agreement. In both experiments, the AI read panoramic images more accurately than OMFS specialists, demonstrating higher diagnostic performance. Table 2 Comparison of diagnostic performance across experiments Experiment 1, true contact position between M3 and IAN Reader Accuracy (%) Sensitivity (%) Specificity (%) Cohen’s kappa Kappa index A 58.04 87.69 17.02 0.05 Slight B 58.04 66.15 46.81 0.13 Slight C 61.61 50.77 76.60 0.26 Fair D 55.36 41.54 74.47 0.15 Slight E 52.68 24.62 91.49 0.12 Slight F 69.64 86.15 46.81 0.35 Fair AI 72.32 84.62 55.32 0.41 Moderate Experiment 2, bucco-lingual position between M3 and IAN A 41.94 60.00 25.00 -0.15 Poor B 38.71 46.67 31.25 -0.22 Poor C 45.16 53.33 37.50 -0.09 Poor D 48.39 100.00 0.00 incalculable incalculable E 51.61 46.67 56.25 0.03 Slight F 32.26 40.00 25.00 -0.35 Poor AI 80.65 86.67 75.00 0.61 Substantial M3, mandibular third molar; IAN, inferior alveolar nerve. Discussion This study evaluated if AI could determine the positional relationship between M3 and IAN based on panoramic radiography regarding whether the two structures were in contact or intimate and whether the IAN was positioned lingually or buccally to M3. AI could determine both positions more accurately than OMFS specialists. Until now, if M3 and IAN overlap on panoramic radiograph, specialists could use the known predictive signs of IAN injury to determine the positional relationship whether the two structures were in contact or intimate. Umar et al. compared the positional relationship between IAN and M3 through panoramic radiography and CBCT. Loss of the radiopaque line and diversion of the canal on panoramic radiographs resulted in tooth and nerve contact in 100% of the cases on CBCT. Darkening of the roots were associated with contact on CBCT in 76.9% of the cases studied 22 . However, another study reported that the sensitivities and specificities ranged from 14.6–68.3% and from 85.5–96.9%, respectively, for those three predictive signs 1 . Datta et al. compared those signs with the clinical findings during surgical removal and found that only 12% of patients with positive radiological signs showed clinical evidence of involvement 3 . In the present study, we adopted CBCT reading results instead of radiological signs on panoramic radiography to determine the positional relationship so that the AI could determine whether the two structures were in contact or intimate, showing an accuracy of 0.55 to 0.72. Compared to another study 1 , our deep learning model exhibited similar performance (accuracy 0.87, precision 0.90, recall, 0.96, F1 score 0.93, and AUC 0.82) to determine whether M3 is contacting the IAN or not. This could explain the different model performance depending on the characteristics of the data. To replace CBCT with analysis of panoramas with AI, information about bucco-lingual positioning was necessary to ensure safe surgical outcomes. It has been reported that the lingual position of the nerve to the tooth has a significantly higher risk of IAN injury compared to other positions 23 . However, there have been few studies reporting the bucco-lingual relationship using plain radiographs. The vertical tube shift technique is a diagnostic method evaluating the bucco-lingual relationship. Nevertheless, this technique caused patient discomfort and nausea during placement of the film or sensor of the digital intraoral x-ray devices 24 and is difficult to use clinically. Since there was no effective method to discern the position, the accuracy of the specialists was low in this study. On the contrary, the AI showed considerably high accuracy ranges from 67.7–80.6% despite the small amount of study data. The course of the IAN predominantly is buccal to the tooth 23 , and our data revealed a similar situation. However, the total number of cases was small to match the numbers in each group evenly for deep learning. Therefore, training AI with more data could produce more accurate results and be used more widely in clinical settings. In this study, bucco-lingual determination (Experiment 2) exhibited superior performance for true contact positioning (Experiment 1). The difference in accuracy between the two experiments seems to be a characteristic of the data rather than a special technical difference. There might be a particular advantage for AI to be recognized in bucco-lingual classification, or that some of the contact classification data might have characteristics that are difficult to distinguish. Panoramic radiography is the most widely used screening test, but image distortion and low resolution of the panoramic images requires possible future examinations with CBCT. It is widely known that CBCT is necessary to confirm subtle changes of the cortical surface. However, the same-side-lingual opposite-side-buccal (SLOB) technique is accepted to determine bucco-lingual positioning 25 , and it might indicate that modalities with better resolution are not required to evaluate the bucco-lingual relationship. If AI could determine the bucco-lingual relationship between M3 and IAN, it could prove to be very helpful before surgery. There are several studies that have developed Al algorithms that have been able to outmatch specialists in terms of performance and accuracy. AI assistance improved the performance of radiologists in distinguishing coronavirus disease 2019 from pneumonia of other origins in chest CT 26 . Moreover, the AI system outperformed radiologists in clinically relevant tasks of breast cancer identification on mammography 27 . In the present study, the AI exhibited much higher accuracy and performance compared to those of OMFS specialists. To determine the positional relationship between M3 and IAN, we performed preliminary tests to determine the most suitable AI model using VGG19, DenseNet, EfficientNet, and ResNet-50. ResNet showed higher AUC in Experiment 2 and comparable AUC in Experiment 1 (Supplemental Tables 1, 2, and 3). Therefore, it was chosen as the final AI model. This study has limitations. First, the absolute size of the training dataset was small. Data augmentation by image modification was used to overcome the limitation of a small sized dataset. Nevertheless, as shown in Table 1 , there were cases where training did not proceed robustly. Therefore, the performances of the trained models highly depend on the train-test split. This unsoundness of the trained model, which hinders the clinical utility of AI models for primary determination in practice, can be alleviated by collecting more data and using them for training. In addition, this study is meaningful in that the AI model performed better than experts even under these adverse conditions. Second, there was no external dataset from multiple dental centers for reproducibility in general utility. We plan to study large datasets including internal and external data to overcome limitations in future studies. Conclusions In this study, we developed and validated a deep learning algorithm that determined positional relationship between M3 and IAN canal at a performance level superior to that of experts. Once tested prospectively in clinical settings, the algorithm could have the potential to narrow patient access to CBCT or prepare for surgical extraction. Declarations Author Contributions E.C. contributed to the conception, design, data acquisition, analysis, and interpretation, and drafted and critically revised the manuscript. K.P. contributed to the conception, design, data acquisition and interpretation, and drafted and critically revised the manuscript. Y.S. contributed to the design, data analysis, and interpretation, and critically revised the manuscript. S.L., E.J., and S.S. contributed to data analysis, and drafted the manuscript. S.Y. contributed to the design, data interpretation, and critically revised the manuscript. All authors gave final approval and agree to be accountable for all aspects of the work. Competing interests The authors declare no competing interests. References Szalma, J., Lempel, E., Jeges, S., Szabo, G. & Olasz, L. The prognostic value of panoramic radiography of inferior alveolar nerve damage after mandibular third molar removal: Retrospective study of 400 cases. Oral Surg. Oral Med. Oral Pathol. Oral Radiol. Endod. 109 , 294-302 (2010). Wang, D. et al. Radiographic features of anatomic relationship between impacted third molar and inferior alveolar canal on coronal cbct images: Risk factors for nerve injury after tooth extraction. Arch. Med. Sci . 14 , 532-540 (2018). Datta, S. et al. Correlation of panoramic radiological and intra-operative findings of impacted mandibular 3rd molar in relation to inferior alveolar canal: A prospective study. J. Maxillofac. Oral. Surg. 1-7 (2020). Fukuda, M. et al. Comparison of 3 deep learning neural networks for classifying the relationship between the mandibular third molar and the mandibular canal on panoramic radiographs. Oral Surg. Oral Med. Oral Pathol. Oral Radiol. 130 , 336-343 (2020). Bell, G.W. Use of dental panoramic tomographs to predict the relation between mandibular third molar teeth and the inferior alveolar nerve. Br. J. Oral. Maxillofac. Surg. 42, 21-27 (2004). Khanagar, S.B. et al . Developments, application, and performance of artificial intelligence in dentistry–a systematic review. J. Dent. Sci. 16 , 508-522 (2021). De Tobel, J., Radesh, P., Vandermeulen, D. & Thevissen, P.W. An automated technique to stage lower third molar development on panoramic radiographs for age estimation: A pilot study. J. Forensic. Odontostomatol. 35 , 42 (2017). Lee, J-S. et al. Osteoporosis detection in panoramic radiographs using a deep convolutional neural network-based computer-assisted diagnosis system: A preliminary study. Dentomaxillofac. Radiol. 48 , 20170344 (2019). Lee, K-S. et al. Evaluation of transfer learning with deep convolutional neural networks for screening osteoporosis in dental panoramic radiographs. J. Clin. Med. 9 , 392 (2020). Fukuda, M. et al. Evaluation of an artificial intelligence system for detecting vertical root fracture on panoramic radiography. Oral Radiol. 1-7 (2019). Tuzoff, D.V. et al. Tooth detection and numbering in panoramic radiographs using convolutional neural networks. Dentomaxillofac. Radiol. 48 , 20180051 (2019). Ekert, T. et al. Deep learning for the radiographic detection of apical lesions. J. Endod. 45 , 917-922. e915 (2019). Murata, M. et al. Deep-learning classification using convolutional neural network for evaluation of maxillary sinusitis on panoramic radiography. Oral Radiol. 35 , 301-307 (2019). Vinayahalingam, S. et al. Automated detection of third molars and mandibular nerve by deep learning. Sci. Rep . 9 , 1-7 (2019). Krois, J. et al. Deep learning for the radiographic detection of periodontal bone loss. Sci. Rep. 9 , 1-6 (2019). Patil, V. et al. Artificial neural network for gender determination using mandibular morphometric parameters: A comparative retrospective study. Cogent. Engineering. 7 , 1723783 (2020). Choi, E., Kim, D., Lee, J-Y. & Park, H-K. Artificial intelligence in detecting temporomandibular joint osteoarthritis on orthopantomogram. Sci. Rep. 11 , 1-7 (2021). Kim, D. et al. Expert system for mandibular condyle detection and osteoarthritis classification in panoramic imaging using r-cnn and cnn. App. Sci. 10 , 7464 (2020). Serte, S., Serener, A. & Al‐Turjman, F. Deep learning in medical imaging: A brief review. Transactions on Emerging Telecommunications Technologies. e4080 (2020). He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition. p. 770-778 (2016). Ajit, A., Acharya, K. & Samanta, A. A review of convolutional neural networks. International Conference on Emerging Trends in Information Technology and Engineering (ic-ETITE). p. 1-5 (2020). Umar, G., Bryant, C., Obisesan, O. & Rood, J. P. Correlation of the radiological predictive factors of inferior alveolar nerve injury with cone beam computed tomography findings. Oral Surg. 3 , 72-82 (2010). Hasegawa, T. et al. Risk factors associated with inferior alveolar nerve injury after extraction of the mandibular third molar—a comparative study of preoperative images by panoramic radiography and computed tomography. Int. J. Oral Maxillofac. Surg. 42 , 843-851 (2013). Jaju, P. P. Localization of mandibular canal by buccal object rule. Oral Surg. Oral Med. Oral Pathol. Oral Radiol. Endod. 109 , 799-author (2010). Rajathi, J., Austin, R., Ramasamy, S. & Mathew, P. Validation of dental panoramic radiographs for the localization of impacted maxillary canines. Indian J. Oral Sci. 5 , 10 (2014). Bai, H.X. et al. Artificial intelligence augmentation of radiologist performance in distinguishing covid-19 from pneumonia of other origin at chest ct. Radiology. 296 , E156-165 (2020). McKinney, S.M. et al. International evaluation of an ai system for breast cancer screening. Nature. 577 , 89-94 (2020). Additional Declarations No competing interests reported. Supplementary Files SciSupp.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Major revision 24 Nov, 2021 Reviews received at journal 28 Oct, 2021 Reviewers agreed at journal 26 Oct, 2021 Reviewers invited by journal 26 Oct, 2021 Editor assigned by journal 26 Oct, 2021 Editor invited by journal 15 Sep, 2021 Submission checks completed at journal 15 Sep, 2021 First submitted to journal 08 Sep, 2021 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-888103","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":52167833,"identity":"a892367d-942a-4706-a19c-85104b4e436c","order_by":0,"name":"Eunhye Choi","email":"","orcid":"","institution":"Seoul National University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Eunhye","middleName":"","lastName":"Choi","suffix":""},{"id":52167834,"identity":"2bfcb586-0b7f-4ef5-b848-388bd0f54a0b","order_by":1,"name":"Soohong Lee","email":"","orcid":"","institution":"Dongguk University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Soohong","middleName":"","lastName":"Lee","suffix":""},{"id":52167835,"identity":"44f61124-727b-4cc7-b7aa-9b85bbd49846","order_by":2,"name":"Eunjae Jeong","email":"","orcid":"","institution":"Dongguk University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Eunjae","middleName":"","lastName":"Jeong","suffix":""},{"id":52167836,"identity":"5887bdbe-ee1a-4972-ba2d-ef1fbbf5d7c3","order_by":3,"name":"Seokwon Shin","email":"","orcid":"","institution":"Dongguk University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Seokwon","middleName":"","lastName":"Shin","suffix":""},{"id":52167837,"identity":"cbae6be4-5ec6-4021-ae1e-ae97c9cf77ed","order_by":4,"name":"Sekyoung Youm","email":"","orcid":"","institution":"Dongguk University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Sekyoung","middleName":"","lastName":"Youm","suffix":""},{"id":52167838,"identity":"7aada042-a27e-488e-9c8a-f0de295f17e7","order_by":5,"name":"Youngdoo Son","email":"","orcid":"","institution":"Dongguk University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Youngdoo","middleName":"","lastName":"Son","suffix":""},{"id":52167839,"identity":"6ed68a31-b0ac-4da9-953a-2d00a43af897","order_by":6,"name":"KangMi Pang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAuElEQVRIiWNgGAWjYDACHgbGBwkVMN4B4rQwGzw4Q6IWNsmHbaRoMec5nWyQOK8uz+AA88MPDGfuEdZi2du78UHitsPFBgfYjCUYbhQT1mJwnnezQeK2A4kbDjCYMTB8SCBKyzaJxDl1QC3s34jUcrYXqKWBGaiFB2jLDSK0WPac3WyQcOxw4szDPMUSCWeI0GLOk7vx4Y+ausS+4+0bP3w4RozD4CxmICZCA7KWUTAKRsEoGAW4AAAdmz96+kFeJAAAAABJRU5ErkJggg==","orcid":"","institution":"Seoul National University Dental Hospital","correspondingAuthor":true,"submittingAuthor":false,"prefix":"","firstName":"KangMi","middleName":"","lastName":"Pang","suffix":""}],"badges":[],"createdAt":"2021-09-09 02:14:05","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-888103/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-888103/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":13783970,"identity":"92662584-bbb6-4b71-8bed-48b187cad1b9","added_by":"auto","created_at":"2021-09-20 16:01:24","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":125866,"visible":true,"origin":"","legend":"Classification of panoramic images based on CBCT. The M3 and IAN seemed to be superimposed in four panoramic images. White triangles point to the border of the IAN in CBCT. \n(A) Intimate but non-contact positioning between M3 and IAN.\n(B) True contact positioning between M3 and IAN.\n(C) IAN positioned buccal to M3.\n(D) IAN positioned lingual to M3.\nCBCT, cone-beam computerized tomography; M3, mandibular third molar; IAN, inferior alveolar nerve. ","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-888103/v1/eb06c057495e1726b73cad54.jpg"},{"id":13783972,"identity":"2d508a80-08f5-4350-aab7-4fe03e0cc566","added_by":"auto","created_at":"2021-09-20 16:01:24","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":107195,"visible":true,"origin":"","legend":"Comparison of sensitivities and specificities of six OMFS specialists and the AI model for determination of the positional relationship between M3 and IAN.\n(A) Experiment 1: Determination of true contact positioning between M3 and IAN. \n(B) Experiment 2: Determination of bucco-lingual positioning between M3 and IAN.\nOMFS, oral and maxillofacial surgery; AI, artificial intelligence; M3, mandibular third molar; IAN, inferior alveolar nerve.","description":"","filename":"Figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-888103/v1/cf011f86785b47a0083131a9.jpg"},{"id":13784285,"identity":"58e6c126-6e96-43bb-bf38-33ad0738dd4b","added_by":"auto","created_at":"2021-09-20 16:04:31","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":446077,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-888103/v1/ac767420-a9d4-4248-a900-8d2f49ae6b1a.pdf"},{"id":13784272,"identity":"1ccd2551-8a22-47d9-b552-2804424e5de6","added_by":"auto","created_at":"2021-09-20 16:04:24","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":18043,"visible":true,"origin":"","legend":"","description":"","filename":"SciSupp.docx","url":"https://assets-eu.researchsquare.com/files/rs-888103/v1/9954b0be64df83215ba70b23.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003eArtificial Intelligence in Positioning Between Tooth and Nerve on Panoramic Radiography\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eMandibular third molar (M3) extraction is one of the most frequently performed surgical procedures in oral and maxillofacial surgery (OMFS). Among the complications following surgery, damage to the inferior alveolar nerve (IAN) is one of the most distressing, causing temporary or permanent neurosensory impairments in the lower lip and chin area at an incidence of 0.4\u0026ndash;13.4%\u003csup\u003e1,2\u003c/sup\u003e. To avoid IAN damage, preoperative assessment of the position of the IAN in relation to the tooth is necessary. Panoramic radiography is used commonly to assess the relationship between M3 and IAN. Certain radiographic features such as darkening of the root and narrowing of the mandibular canal have been reported as risk factors for IAN injuries, although its clinical correlation was low\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. Due to the development of cone-beam computerized tomography (CBCT), determination of positioning between the IAN and teeth has become more accurate, and CBCT is recommended before M3 extraction when the two aforementioned structures are superimposed on panoramic radiography\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. However, the disadvantages of CBCT include higher radiation doses compared to two-dimensional imaging and the presence of image artifacts mainly produced by metal restorations\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eTherefore, accurate methods diagnosing the relationship between M3 and IAN on panoramic radiography are necessary. After the diagnostic methods determine whether both structures are truly in contact or intimate, assessment whether M3 is positioned lingually or buccally to the IAN is necessary to determine the direction of insertion of the surgical instruments.\u003c/p\u003e\n\u003cp\u003eArtificial intelligence (AI) models have reported excellent performance, mimicking the precision and accuracy of trained specialists in dentistry\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. Various studies have applied AI algorithms to read panoramic radiographs for clinical conditions such as age estimation\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e, osteoporosis\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e8\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e, vertical root fracture\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e, automatic teeth detection and numbering\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e, apical lesions\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e, maxillary sinusitis\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e, detecting and segmenting the approximation of the inferior alveolar nerve and mandibular third molar\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e, periodontal bone loss\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e, gender determination\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e, and temporomandibular joint osteoarthritis\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e17\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. However, there is an AI study on the positional relationship between M3 and IAN, focusing on contact or non-contact\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eThis study aimed to investigate the clinical use of an AI model developed to determinate the positional relationship between M3 and IAN from panoramic radiography using deep learning that compared the AI readings with those of OMFS specialists.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003eMaterials.\u003c/strong\u003e The written documentation of informed consent was waived and approved by the decision of the Institutional Review Board of Seoul National University Dental Hospital (ERI21004) and ethics committee approval for the study in the same institute was also obtained. All methods were performed in accordance with the relevant guidelines and regulation. Subjects were included retrospectively from an image database of patients who visited the Department of Oral and Maxillofacial Surgery at Seoul National University Gwanak Dental Hospital between January 2019 and December 2020. Patients who underwent both panoramic radiography (Kodak 8000 Digital Panoramic System, Trophy Radiologies, Carestream Health Inc., NY, USA) and CBCT (CS 9300, Carestream Health Inc., NY, USA) for M3 extraction with superimposition of M3 and IAN on the panoramic radiographs were selected. The patients consisted of 200 males and 194 females, with an age range of 20 to 72 years (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD age, 31.5 4\u0026thinsp;\u0026plusmn;\u0026thinsp;9.96 y; range, 20 to 72 y). The panoramic images of 571 M3s from these patients were used in this study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAI model developments.\u003c/strong\u003e The AI model was developed to evaluate two positional relationships between M3 and IAN.\u003c/p\u003e\n\u003col\u003e\n\u003cli\u003e\n\u003cp\u003eExperiment 1: Determination of the true contact position between M3 and IAN\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eExperiment 2: Determination of the bucco-lingual position between M3 and IAN\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003ePanoramic images that appeared overlapped were classified as true contact and intimate according to the presence or absence of the cortical line of the IAN canal on CBCT (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003eA, B). Independently, the bucco-lingual positional relationship was also confirmed by CBCT (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003eC, D). Determination of the positional relationship based on the CBCT was performed by an OMFR specialist. Among the 571 images, 202 were classified as true contact, 246 as false contact, 61 as IAN buccal position, and 62 as IAN lingual position. Regions of Interest (ROI) were extracted from the panoramic radiograph manually in JPG format with a matrix size of 400 x 400 pixels.\u003c/p\u003e\n\u003cp\u003eResNet-50, mainly used for medical image classification\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e, is a substantially deeper and easier model to train compared to simple models such as VGGnet, and the core structure is a residual block\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. Residual learning does not allow for error accumulation on the convolution layers but enables a better representation of the content in the convolution layers. By adopting a shortcut structure, the vanishing gradient issue is resolved\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. Every image was resized to 224 x 224 pixels, and we randomly split the dataset into 75% for training and validation and 25% for testing. The model performance varies depending on the difficulty of data classification, so we performed 5 repeated experiments through random sampling. As a technical and strategic method to avoid overfitting, data augmentation was performed by image rotation\u0026thinsp;\u0026plusmn;\u0026thinsp;30 degrees, horizontal flipping, and brightness 20\u0026ndash;80% for every mini-batch in training to compensate for the small number of data points to increase model robustness. In Experiment 1, a model was trained for 60 epochs with augmented data. The learning rate of the model was set to 1.0 \u0026times; 10\u0026ndash;4 and an Adam optimizer was used. In Experiment 2, training was progressed in 30 epochs with augmented data. In addition, the learning rate of the model was 1.0 \u0026times; 10\u0026ndash;4, and an Adam optimizer was used.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel and statistical analysis.\u003c/strong\u003e Accuracy, precision, recall, F1 score, and AUC were calculated to evaluate each model performance. Accuracy is defined as the ratio of correct predictions. Precision is the ratio of true positives to true positives and false positives. Recall is the ratio of true positives to true positives and false negatives. F1 score is a harmonic mean of precision and recall: (2 \u0026times; precision \u0026times; recall)/(precision\u0026thinsp;+\u0026thinsp;recall), and AUC is the area under the ROC curve. The confidence intervals of AUC were found by bootstrapping with 1,000 test sets sampled with replacement. For evaluation of AI clinical usability, the results between OPG reads by the AI and six OMFS specialists were compared. Accuracy, sensitivity, and specificity were calculated for diagnostic performance, and Cohen\u0026rsquo;s kappa was calculated to estimate the strength of agreement. Among the random samplings, the same test dataset with the highest accuracy was selected for comparison with specialists in each experiment. Python programming language (v. 3.8.5), Tensorflow (v. 2.5.0), and a graphics card (GeForce RTX 3090) were used for analysis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data that support the findings this study are available from the authors upon reasonable request.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eTable \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e shows the model performance in each experiment. The average accuracy, precision, recall, and F1 score were 0.63, 0.62, 0.63, and 0.61, respectively, in Experiment 1, with true contact position determination between M3 and IAN. The average accuracy, precision, recall, and F1 score were 0.76, 0.83, 0.67, and 0.73, respectively, in Experiment 2, with bucco-lingual position determination between M3 and IAN. The overall model performance was superior in Experiment 2 compared to Experiment 1.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\u003ctable border=\"1\" id=\"Tab1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eModel performance of five random samplings in each experiment\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"6\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" colspan=\"6\"\u003e\n \u003cp\u003eExperiment 1, true contact position between M3 and IAN\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eWork\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1 score\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUC\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.72\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.72\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.75\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.74\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.62\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.62\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.67\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.68\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.66\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.62\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.57\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.64\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAverage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.62\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.66\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"6\"\u003e\n \u003cp\u003e\u003cstrong\u003eExperiment 2, bucco-lingual position between M3 and IAN\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.77\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.86\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.74\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.77\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.75\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.77\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.57\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.79\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.68\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.47\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.62\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAverage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.76\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.73\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003ctfoot\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"6\"\u003eM3, mandibular third molar; IAN, inferior alveolar nerve; AUC, Area under the ROC curve.\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tfoot\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n\u003cp\u003eThe comparison of sensitivities and specificities between AI and OMFS specialists in each experiment is shown in Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e. AI exhibited 72.32% accuracy in Experiment 1 and 80.65% in Experiment 2, but the highest accuracy among OMFR specialists in each experiment was 69.64% and 51.61%, respectively (Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e). Cohen\u0026rsquo;s kappa of AI was highest in Experiment 2 and showed a substantial level of agreement (0.61), but those of OMFS specialists exhibited a slight to fair level of agreement. In both experiments, the AI read panoramic images more accurately than OMFS specialists, demonstrating higher diagnostic performance.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\u003ctable border=\"1\" id=\"Tab2\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eComparison of diagnostic performance across experiments\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"10\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" colspan=\"10\"\u003e\n \u003cp\u003eExperiment 1, true contact position between M3 and IAN\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003eReader\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003eSensitivity (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecificity (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eCohen\u0026rsquo;s kappa\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eKappa index\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e58.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e87.69\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e17.02\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSlight\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eB\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e58.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e66.15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e46.81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSlight\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e61.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e50.77\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e76.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFair\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eD\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e55.36\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e41.54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e74.47\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSlight\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e52.68\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e24.62\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e91.49\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSlight\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e69.64\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e86.15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e46.81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.35\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFair\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eAI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e72.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e84.62\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e55.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.41\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eModerate\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"10\"\u003e\n \u003cp\u003e\u003cstrong\u003eExperiment 2, bucco-lingual position between M3 and IAN\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e41.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e60.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e25.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e-0.15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePoor\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eB\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e38.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e46.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e31.25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e-0.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePoor\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e45.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e53.33\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e37.50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e-0.09\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePoor\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eD\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e48.39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e100.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eincalculable\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eincalculable\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e51.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e46.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e56.25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e0.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSlight\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e32.26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e40.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e25.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e-0.35\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePoor\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e80.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e86.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e75.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003e0.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSubstantial\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003ctfoot\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"10\"\u003eM3, mandibular third molar; IAN, inferior alveolar nerve.\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tfoot\u003e\n \u003c/table\u003e\n\u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study evaluated if AI could determine the positional relationship between M3 and IAN based on panoramic radiography regarding whether the two structures were in contact or intimate and whether the IAN was positioned lingually or buccally to M3. AI could determine both positions more accurately than OMFS specialists.\u003c/p\u003e \u003cp\u003eUntil now, if M3 and IAN overlap on panoramic radiograph, specialists could use the known predictive signs of IAN injury to determine the positional relationship whether the two structures were in contact or intimate. Umar et al. compared the positional relationship between IAN and M3 through panoramic radiography and CBCT. Loss of the radiopaque line and diversion of the canal on panoramic radiographs resulted in tooth and nerve contact in 100% of the cases on CBCT. Darkening of the roots were associated with contact on CBCT in 76.9% of the cases studied\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. However, another study reported that the sensitivities and specificities ranged from 14.6\u0026ndash;68.3% and from 85.5\u0026ndash;96.9%, respectively, for those three predictive signs\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Datta et al. compared those signs with the clinical findings during surgical removal and found that only 12% of patients with positive radiological signs showed clinical evidence of involvement\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. In the present study, we adopted CBCT reading results instead of radiological signs on panoramic radiography to determine the positional relationship so that the AI could determine whether the two structures were in contact or intimate, showing an accuracy of 0.55 to 0.72. Compared to another study\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e, our deep learning model exhibited similar performance (accuracy 0.87, precision 0.90, recall, 0.96, F1 score 0.93, and AUC 0.82) to determine whether M3 is contacting the IAN or not. This could explain the different model performance depending on the characteristics of the data.\u003c/p\u003e \u003cp\u003eTo replace CBCT with analysis of panoramas with AI, information about bucco-lingual positioning was necessary to ensure safe surgical outcomes. It has been reported that the lingual position of the nerve to the tooth has a significantly higher risk of IAN injury compared to other positions\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. However, there have been few studies reporting the bucco-lingual relationship using plain radiographs. The vertical tube shift technique is a diagnostic method evaluating the bucco-lingual relationship. Nevertheless, this technique caused patient discomfort and nausea during placement of the film or sensor of the digital intraoral x-ray devices\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e and is difficult to use clinically. Since there was no effective method to discern the position, the accuracy of the specialists was low in this study. On the contrary, the AI showed considerably high accuracy ranges from 67.7\u0026ndash;80.6% despite the small amount of study data. The course of the IAN predominantly is buccal to the tooth\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e, and our data revealed a similar situation. However, the total number of cases was small to match the numbers in each group evenly for deep learning. Therefore, training AI with more data could produce more accurate results and be used more widely in clinical settings.\u003c/p\u003e \u003cp\u003eIn this study, bucco-lingual determination (Experiment 2) exhibited superior performance for true contact positioning (Experiment 1). The difference in accuracy between the two experiments seems to be a characteristic of the data rather than a special technical difference. There might be a particular advantage for AI to be recognized in bucco-lingual classification, or that some of the contact classification data might have characteristics that are difficult to distinguish.\u003c/p\u003e \u003cp\u003ePanoramic radiography is the most widely used screening test, but image distortion and low resolution of the panoramic images requires possible future examinations with CBCT. It is widely known that CBCT is necessary to confirm subtle changes of the cortical surface. However, the same-side-lingual opposite-side-buccal (SLOB) technique is accepted to determine bucco-lingual positioning\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e, and it might indicate that modalities with better resolution are not required to evaluate the bucco-lingual relationship. If AI could determine the bucco-lingual relationship between M3 and IAN, it could prove to be very helpful before surgery.\u003c/p\u003e \u003cp\u003eThere are several studies that have developed Al algorithms that have been able to outmatch specialists in terms of performance and accuracy. AI assistance improved the performance of radiologists in distinguishing coronavirus disease 2019 from pneumonia of other origins in chest CT\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. Moreover, the AI system outperformed radiologists in clinically relevant tasks of breast cancer identification on mammography\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e. In the present study, the AI exhibited much higher accuracy and performance compared to those of OMFS specialists. To determine the positional relationship between M3 and IAN, we performed preliminary tests to determine the most suitable AI model using VGG19, DenseNet, EfficientNet, and ResNet-50. ResNet showed higher AUC in Experiment 2 and comparable AUC in Experiment 1 (Supplemental Tables\u0026nbsp;1, 2, and 3). Therefore, it was chosen as the final AI model.\u003c/p\u003e \u003cp\u003eThis study has limitations. First, the absolute size of the training dataset was small. Data augmentation by image modification was used to overcome the limitation of a small sized dataset. Nevertheless, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, there were cases where training did not proceed robustly. Therefore, the performances of the trained models highly depend on the train-test split. This unsoundness of the trained model, which hinders the clinical utility of AI models for primary determination in practice, can be alleviated by collecting more data and using them for training. In addition, this study is meaningful in that the AI model performed better than experts even under these adverse conditions. Second, there was no external dataset from multiple dental centers for reproducibility in general utility. We plan to study large datasets including internal and external data to overcome limitations in future studies.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eIn this study, we developed and validated a deep learning algorithm that determined positional relationship between M3 and IAN canal at a performance level superior to that of experts. Once tested prospectively in clinical settings, the algorithm could have the potential to narrow patient access to CBCT or prepare for surgical extraction.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eE.C. contributed to the conception, design, data acquisition, analysis, and interpretation, and drafted and critically revised the manuscript. K.P. contributed to the conception, design, data acquisition and interpretation, and drafted and critically revised the manuscript. Y.S. contributed to the design, data analysis, and interpretation, and critically revised the manuscript. S.L., E.J., and S.S. contributed to data analysis, and drafted the manuscript. S.Y. contributed to the design, data interpretation, and critically revised the manuscript. All authors gave final approval and agree to be accountable for all aspects of the work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003c/strong\u003eThe authors declare no competing interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eSzalma, J., Lempel, E., Jeges, S., Szabo, G. \u0026amp; Olasz, L. The prognostic value of panoramic radiography of inferior alveolar nerve damage after mandibular third molar removal: Retrospective study of 400 cases. \u003cem\u003eOral Surg. Oral Med. Oral Pathol. Oral Radiol. Endod.\u003c/em\u003e \u003cstrong\u003e109\u003c/strong\u003e, 294-302 (2010).\u003c/li\u003e\n \u003cli\u003eWang, D. \u003cem\u003eet al.\u003c/em\u003e Radiographic features of anatomic relationship between impacted third molar and inferior alveolar canal on coronal cbct images: Risk factors for nerve injury after tooth extraction. \u003cem\u003eArch. Med. Sci\u003c/em\u003e. \u003cstrong\u003e14\u003c/strong\u003e, 532-540 (2018).\u003c/li\u003e\n \u003cli\u003eDatta, S. \u003cem\u003eet al.\u003c/em\u003e Correlation of panoramic radiological and intra-operative findings of impacted mandibular 3rd molar in relation to inferior alveolar canal: A prospective study. \u003cem\u003eJ. Maxillofac. Oral. Surg.\u0026nbsp;\u003c/em\u003e1-7 (2020).\u003c/li\u003e\n \u003cli\u003eFukuda, M. \u003cem\u003eet al.\u003c/em\u003e Comparison of 3 deep learning neural networks for classifying the relationship between the mandibular third molar and the mandibular canal on panoramic radiographs. \u003cem\u003eOral Surg. Oral Med. Oral Pathol. Oral Radiol.\u003c/em\u003e \u003cstrong\u003e130\u003c/strong\u003e, 336-343 (2020).\u003c/li\u003e\n \u003cli\u003eBell, G.W. Use of dental panoramic tomographs to predict the relation between mandibular third molar teeth and the inferior alveolar nerve. \u003cem\u003eBr. J. Oral. Maxillofac. Surg.\u003c/em\u003e 42, 21-27 (2004).\u003c/li\u003e\n \u003cli\u003eKhanagar, S.B. \u003cem\u003eet al\u003c/em\u003e. Developments, application, and performance of artificial intelligence in dentistry\u0026ndash;a systematic review. \u003cem\u003eJ. Dent. Sci.\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 508-522 (2021).\u003c/li\u003e\n \u003cli\u003eDe Tobel, J., Radesh, P., Vandermeulen, D. \u0026amp; Thevissen, P.W. An automated technique to stage lower third molar development on panoramic radiographs for age estimation: A pilot study. \u003cem\u003eJ. Forensic. Odontostomatol.\u003c/em\u003e \u003cstrong\u003e35\u003c/strong\u003e, 42 (2017).\u003c/li\u003e\n \u003cli\u003eLee, J-S. \u003cem\u003eet al.\u003c/em\u003e Osteoporosis detection in panoramic radiographs using a deep convolutional neural network-based computer-assisted diagnosis system: A preliminary study. \u003cem\u003eDentomaxillofac. Radiol.\u003c/em\u003e \u003cstrong\u003e48\u003c/strong\u003e, 20170344 (2019).\u003c/li\u003e\n \u003cli\u003eLee, K-S. \u003cem\u003eet al.\u003c/em\u003e Evaluation of transfer learning with deep convolutional neural networks for screening osteoporosis in dental panoramic radiographs. \u003cem\u003eJ. Clin. Med.\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 392 (2020).\u003c/li\u003e\n \u003cli\u003eFukuda, M. \u003cem\u003eet al.\u003c/em\u003e Evaluation of an artificial intelligence system for detecting vertical root fracture on panoramic radiography. \u003cem\u003eOral Radiol.\u003c/em\u003e 1-7 (2019).\u003c/li\u003e\n \u003cli\u003eTuzoff, D.V. \u003cem\u003eet al.\u003c/em\u003e Tooth detection and numbering in panoramic radiographs using convolutional neural networks. \u003cem\u003eDentomaxillofac. Radiol.\u003c/em\u003e \u003cstrong\u003e48\u003c/strong\u003e, 20180051 (2019).\u003c/li\u003e\n \u003cli\u003eEkert, T. \u003cem\u003eet al.\u003c/em\u003e Deep learning for the radiographic detection of apical lesions. \u003cem\u003eJ. Endod.\u003c/em\u003e \u003cstrong\u003e45\u003c/strong\u003e, 917-922. e915 (2019).\u003c/li\u003e\n \u003cli\u003eMurata, M. \u003cem\u003eet al.\u0026nbsp;\u003c/em\u003eDeep-learning classification using convolutional neural network for evaluation of maxillary sinusitis on panoramic radiography. \u003cem\u003eOral Radiol.\u003c/em\u003e \u003cstrong\u003e35\u003c/strong\u003e, 301-307 (2019).\u003c/li\u003e\n \u003cli\u003eVinayahalingam, S. \u003cem\u003eet al.\u003c/em\u003e Automated detection of third molars and mandibular nerve by deep learning. \u003cem\u003eSci. Rep\u003c/em\u003e. \u003cstrong\u003e9\u003c/strong\u003e, 1-7 (2019).\u003c/li\u003e\n \u003cli\u003eKrois, J. \u003cem\u003eet al.\u0026nbsp;\u003c/em\u003eDeep learning for the radiographic detection of periodontal bone loss. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 1-6 (2019).\u003c/li\u003e\n \u003cli\u003ePatil, V. \u003cem\u003eet al.\u003c/em\u003e Artificial neural network for gender determination using mandibular morphometric parameters: A comparative retrospective study. \u003cem\u003eCogent. Engineering.\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, 1723783 (2020).\u003c/li\u003e\n \u003cli\u003eChoi, E., Kim, D., Lee, J-Y. \u0026amp; Park, H-K. Artificial intelligence in detecting temporomandibular joint osteoarthritis on orthopantomogram. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 1-7 (2021).\u003c/li\u003e\n \u003cli\u003eKim, D. \u003cem\u003eet al.\u003c/em\u003e Expert system for mandibular condyle detection and osteoarthritis classification in panoramic imaging using r-cnn and cnn. \u003cem\u003eApp. Sci.\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 7464 (2020).\u003c/li\u003e\n \u003cli\u003eSerte, S., Serener, A. \u0026amp; Al‐Turjman, F. Deep learning in medical imaging: A brief review. \u003cem\u003eTransactions on Emerging Telecommunications Technologies.\u0026nbsp;\u003c/em\u003ee4080 (2020).\u003c/li\u003e\n \u003cli\u003eHe, K., Zhang, X., Ren, S. \u0026amp; Sun, J. Deep residual learning for image recognition. \u003cem\u003eProceedings of the IEEE conference on computer vision and pattern recognition.\u003c/em\u003e p. 770-778 (2016).\u003c/li\u003e\n \u003cli\u003eAjit, A., Acharya, K. \u0026amp; Samanta, A. A review of convolutional neural networks. \u003cem\u003eInternational Conference on Emerging Trends in Information Technology and Engineering (ic-ETITE).\u003c/em\u003e p. 1-5 (2020).\u003c/li\u003e\n \u003cli\u003eUmar, G., Bryant, C., Obisesan, O. \u0026amp; Rood, J. P. Correlation of the radiological predictive factors of inferior alveolar nerve injury with cone beam computed tomography findings. \u003cem\u003eOral Surg.\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 72-82 (2010).\u003c/li\u003e\n \u003cli\u003eHasegawa, T. \u003cem\u003eet al.\u003c/em\u003e Risk factors associated with inferior alveolar nerve injury after extraction of the mandibular third molar\u0026mdash;a comparative study of preoperative images by panoramic radiography and computed tomography. \u003cem\u003eInt. J. Oral Maxillofac. Surg.\u003c/em\u003e \u003cstrong\u003e42\u003c/strong\u003e, 843-851 (2013).\u003c/li\u003e\n \u003cli\u003eJaju, P. P. Localization of mandibular canal by buccal object rule. \u003cem\u003eOral Surg. Oral Med. Oral Pathol. Oral Radiol. Endod.\u003c/em\u003e \u003cstrong\u003e109\u003c/strong\u003e, 799-author (2010).\u003c/li\u003e\n \u003cli\u003eRajathi, J., Austin, R., Ramasamy, S. \u0026amp; Mathew, P. Validation of dental panoramic radiographs for the localization of impacted maxillary canines. \u003cem\u003eIndian J. Oral Sci.\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, 10 (2014).\u003c/li\u003e\n \u003cli\u003eBai, H.X. \u003cem\u003eet al.\u003c/em\u003e Artificial intelligence augmentation of radiologist performance in distinguishing covid-19 from pneumonia of other origin at chest ct. \u003cem\u003eRadiology.\u0026nbsp;\u003c/em\u003e\u003cstrong\u003e296\u003c/strong\u003e, E156-165 (2020).\u003c/li\u003e\n \u003cli\u003eMcKinney, S.M. \u003cem\u003eet al.\u003c/em\u003e International evaluation of an ai system for breast cancer screening. \u003cem\u003eNature.\u003c/em\u003e \u003cstrong\u003e577\u003c/strong\u003e, 89-94 (2020).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Artificial intelligence, deep learning, inferior alveolar nerve, mandibular third molar, panoramic radiography","lastPublishedDoi":"10.21203/rs.3.rs-888103/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-888103/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eDetermining the exact positional relationship between mandibular third molar (M3) and inferior alveolar nerve (IAN) is important for surgical extractions. Panoramic radiography is the most common dental imaging test. The purposes of this study were to develop an artificial intelligence (AI) model to determine two positional relationships (true contact and bucco-lingual position) between M3 and IAN when they were overlapped in panoramic radiographs and compare its performance with that of oral and maxillofacial surgery (OMFS) specialists. A total of 571 panoramic images of M3 from 394 patients was used for this study. Among the images, 202 were classified as true contact, 246 as intimate, 61 as IAN buccal position, and 62 as IAN lingual position. A deep convolutional neural network model with ResNet-50 architecture was trained for each task. We randomly split the dataset into 75% for training and validation and 25% for testing. Model performance was superior in bucco-lingual position determination (accuracy 0.76, precision 0.83, recall 0.67, and F1 score 0.73) to true contact position determination (accuracy 0.63, precision 0.62, recall 0.63, and F1 score 0.61). AI exhibited much higher accuracy in both position determinations compared to OMFS specialists. In determining true contact position, OMFS specialists demonstrated an accuracy of 52.68% to 69.64%, while the AI showed an accuracy of 72.32%. In determining bucco-lingual position, OMFS specialists showed an accuracy of 32.26% to 48.39%, and the AI showed an accuracy of 80.65%. Moreover, Cohen’s kappa exhibited a substantial level of agreement for the AI (0.61) and poor agreements for OMFS specialists in bucco-lingual position determination. Determining the position relationship between M3 and IAN is possible using AI, especially in bucco-lingual positioning. The model could be used to support clinicians in the decision-making process for M3 treatment.\u0026nbsp;\u003c/p\u003e","manuscriptTitle":"Artificial Intelligence in Positioning Between Tooth and Nerve on Panoramic Radiography","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2021-09-20 16:01:22","doi":"10.21203/rs.3.rs-888103/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Major revision","date":"2021-11-24T08:30:11+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2021-10-28T07:04:46+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"503bb26b-1aef-49ae-90dc-44ff80b6f2f8","date":"2021-10-26T19:08:27+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2021-10-26T18:06:01+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2021-10-26T12:31:54+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2021-09-15T09:36:21+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2021-09-15T09:29:49+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2021-09-09T02:04:57+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"5537509b-eab2-4b0e-b048-49c8475e7088","owner":[],"postedDate":"September 20th, 2021","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":7311029,"name":"Dentistry"}],"tags":[],"updatedAt":"2022-01-06T08:59:05+00:00","versionOfRecord":[],"versionCreatedAt":"2021-09-20 16:01:22","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-888103","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-888103","identity":"rs-888103","version":["v1"]},"buildId":"FbvkV6FR0MCFSLy54lSbu","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}