Deep Learning-based Diagnosis of Cervical Burnout and Interproximal Caries in Bitewing Radiographs

Deep Learning-based Diagnosis of Cervical Burnout and Interproximal Caries in Bitewing Radiographs | 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 Deep Learning-based Diagnosis of Cervical Burnout and Interproximal Caries in Bitewing Radiographs Ozcan KARATAS, Ridvan AKYOL, Kemal Selcuk YUCEL, Ebru DELIKAN, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6693428/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background: This study aimed to evaluate the success of deep learning-based convolutional neural networks (CNN) and Residual Neural Network-34 (ResNet34) in classifying and detecting cervical burnout and interproximal caries on bitewing radiographs. Methods: Within the scope of the study, a dataset consisting of 454 bitewing radiographs, free of noise and artifacts, was labeled by two dentists with LabelImg software. A 32-layer CNN model (614 interproximal caries, 402 cervical burnout) was created for classification, and a ResNet34 model was created for object detection. The images were resized to 300x300 pixels, and the datasets were divided into 80% training and 20% testing. Performance metrics included sensitivity, specificity, precision, accuracy, and F1 score. Results: The classification model achieved 93.14% accuracy, 86.42% sensitivity, and 97.56% specificity, while the object detection model gave 81.74% accuracy and 0.82 mAP values ​​at 0.5 IoU. The data showed that CNN models were successful in classifying cervical burnout and interproximal caries. The data also showed that ResNet-34 models were successful in detecting cervical burnout and interproximal caries. Conclusion: Despite a limited dataset, CNN models showed successful results in classifying cervical burnout and interproximal caries on bitewing radiographs. Cervical Burnout Convolutional Neural Networks Deep Learning Interproximal Caries Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Background Advancements in artificial intelligence (AI) applications are currently resulting in significant changes in the field of health sciences. [ 1 ] Machine learning (ML) is a highly dynamic and swiftly evolving branch of artificial intelligence predicated on the concept of empowering computers to learn from data. Deep learning, a subset of machine learning, employs multi-layered artificial neural networks to discern intricate data relationships. Deep neural networks, modeled after the functionality of human brain neurons, process data hierarchically to extract advanced features. Unlike conventional image processing and machine learning methods that necessitate manual feature extraction, deep learning models, particularly Convolutional Neural Networks (CNN), autonomously learn these features, enhancing the efficiency of the analysis process. [ 2 , 3 ] Therefore, CNNs are extensively utilized in dentistry owing to their exceptional efficacy, particularly in image-based evaluations. Research indicates that deep learning-based CNNs yield effective outcomes in various domains, including tooth classification from 2D and 3D images, detection of caries and lesions, assessment of anatomical structures, orthodontic evaluations, and identification of restorations. [ 4 ] Traditional manual evaluation methods can be challenging to use and take a long time. Deep learning systems, on the other hand, can quickly and effectively analyze large datasets to help clinicians make decisions. Furthermore, the implementation of such systems enhances the reliability and consistency of diagnoses for patients by minimizing the margin of error in dentistry. [ 5 ] Dental caries are among the most prevalent ailments encountered in society. Early diagnosis and treatment of caries is fundamental to protecting oral and dental health. [ 6 ] Alongside intraoral, visual, and tactile examinations, both traditional and digital radiography are commonly used in the identification of dental caries. Currently, digital radiographs are increasingly favored over traditional radiography owing to numerous advantages, including reduced application time, enhanced storage capabilities, and improved data processing and transfer. [ 7 ] Bite-wing radiographs are among the most frequently utilized diagnostic techniques, particularly for identifying interproximal caries. Radiographs play a crucial role in diagnostic dentistry as they reveal details that are invisible to the human eye, such as interproximal cavities, bone levels, existing restorations, and secondary cavities. [ 8 ] Nonetheless, artifacts seen on radiographs might hinder clinicians' evaluations, make it challenging to evaluate the patient, and may even result in an inaccurate diagnosis. A phenomenon known as cervical burnout is a radiolucent band or wedge-shaped structure seen in the cervical areas of teeth on radiographs. [ 9 ] Cervical burnout, an artifact resulting from the differential absorption of X-rays by radiopaque structures like enamel, dentin, cementum, and bone, is not a genuine pathology. However, it is a concern highlighted that practitioners should not conflate this artifact with interproximal caries. Likewise, burnouts observed beneath interproximal restorations may be confused with secondary caries. [ 10 ] To accurately tell the difference between interproximal caries and cervical burnout, it is suggested that the clinical and radiographic evaluations be done at the same time. In such cases, assessments conducted on radiographic images utilizing supportive AI systems may help clinicians in achieving accurate diagnoses and save time. Therefore, assessing the effectiveness of AI applications on this topic will be an essential issue of discussion. Unfortunately, to the best of our knowledge, no study has been conducted in the literature to investigate the effectiveness of deep learning-based ResNets in detecting cervical burnout and their potential for differential diagnosis with caries. Within this framework, our objective is to assess the effectiveness of ResNets in the diagnosis of cervical burnout and interproximal caries in bitewing radiographs. Since cervical burnout and interproximal caries have similar appearances, problems may occur both in the object detection phase and in the classification phase, and the detected lesion may be classified incorrectly. Especially in the studies of Lee et al.[ 11 ] using the Inceptionv3 model, the False Negative rates primarily led us to the classification task. Therefore, we examined the classification success with the CNN model, which has proven its performance in classification tasks in dental radiographs.[ 12 ] Although there are studies in the literature with CNN-based models such as Mask R-CNN with small data sets, we did not come across a study on the detection algorithm with the ResNet model prepared for small data sets. [ 13 ] Methods Study design and ethical approval This retrospective study was conducted using a dataset consisting of bitewing radiographs available in the archives of the Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, Nuh Naci Yazgan University. The study was approved by the Nuh Naci Yazgan University Non-Interventional Clinical Research Ethics Committee (decision date and number: 2025/03-001), and the study protocol was carried out following the ethical guidelines stated in the Declaration of Helsinki. Dataset Creation and Image Preprocessing Radiographs were randomly selected by two dentists from the archives of the Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, XXX University, containing bitewing radiographs taken between January 2024 and January 2025 for the diagnosis and treatment of caries. Only radiographs containing permanent teeth were used as data, without any additional information about the patients (e.g., gender, age, or other clinical information). Radiographs with poor image quality, excessive noise, artifacts, or severe overlapping of proximal surfaces due to anatomical alignment of certain teeth were excluded. In addition, all radiographic images used in this study were obtained using the same device (KaVo X-ray machine; KaVo FOCUS, Tuusula, Finland) using the irradiation parameters of 60 kV, 7 mA, and 0.08 s. A restorative dentist with at least 10 years of experience and an oral radiologist were able to label the radiograph images together, and images that could not be agreed upon were excluded. Since the aim was to distinguish between interproximal caries and cervical burnouts on bitewing images, it was decided to create both a classification model (Classification) and an object detection model (Detection). Then, only interproximal caries and cervical burnout images were labeled for the object detection model, and LabelImg software was used for labeling (Copyright (C) 2013 MIT, Computer Science and Artificial Intelligence Laboratory) [14] (Figure 1). Secondary caries under restoration and caries with high occlusal loss were excluded from the study to not affect performance. At the same time, the labeled regions of interest (ROI) were cropped and exported for the classification task (Figure 2). All bitewing images were exported in PNG format, while the labels were saved in XML (Extensible Markup Language) format. The final dataset consisted of 454 bitewing radiographs, 614 interproximal caries, and 402 cervical burnout images. Network Training Before training the classification model, preprocessing steps were performed, which included resizing the images to 300 × 300 pixels. The dataset was then split into two groups; for interproximal caries, the training set consisted of 491 images (80%) and the test set consisted of 123 images (20%). For cervical burnouts, the training set consisted of 321 images (80%), and the test set consisted of 81 images (20%). Before training the object detection model, preprocessing steps were performed, which included resizing the images to 600 × 600 pixels. The dataset was then split into three separate groups, the training set consisting of 364 images (80%) and the test set consisting of 90 images (20%) (Table 1). Table1. Dataset and label distribution Dataset Total Bitewing Radiographs 454 Total Interproximal Caries 614 Total Cervical Burnout 402 Total Number of Labels 1016 Model Training Set Test Set Classification Model CNN/ Number of Labels 491 + 321 123 + 81 Classification Model CNN/ Number of Images 491 + 321 123 + 81 Object Detection Task ResNet34/ Number of Labels 491 + 321 123 + 81 Object Detection Task ResNet34 364 90 Deep Learning Procedure For the classification task, a 32-layer CNN model was created with the Keras application from the TensorFlow library. While the activation function “SoftMax” was used in the last layer, “ReLU” was used in the other layers. The parameters “binary_crossentropy,” “rmsprop,” and 32 as batch_size were used as the loss function. The evaluation metrics were carried out using the confusion matrix (Figure 3). A ResNet34 model (Residual Network) was created for the object detection task via PyTorch (Figure 4). With the torchvision.models.resnet34 function, the weights parameter ResNet34_Weights is optional; the process parameter (progress) is a bool, optional, and the kwargs parameters are any; the optimizer parameter is "Adam," the learning rate parameter is "0.001," and the batch size parameter was determined as 64. [15] ResNet34 model (Residual Network) Introduced by Microsoft Research in 2015, ResNet brought a revolutionary approach to deep learning by utilizing residual blocks to facilitate the training of deeper networks. It is a state-of-the-art CNN architecture that enables networks to reach substantial depths while achieving remarkable success in numerous applications. Residual connections are employed to address challenges encountered during the training of deep neural networks, such as the vanishing gradient problem. This approach allows for the construction of more powerful and complex models by adding additional layers. [16] Metrics for the model’s performance The evaluation metrics were determined as sensitivity, specificity, precision, accuracy, and F1 score. Sensitivity (TPR) represents the true positive rate and indicates the likelihood of correctly identifying a disease. TPR = TP / (TP + FN) (TPR: true positive rate, TP: true positive, FN: false negative) Specificity (SPC) represents the true negative rate and indicates the likelihood of correctly identifying healthy individuals. SPC = TN / (FP + TN) (SPC: specificity, TN: true negative, FP: false positive) Precision (PPV) is the positive predictive value and shows the probability that identified positives are truly diseased. PPV = TP / (TP + FP) (PPV: positive predictive value) Accuracy (ACC) represents the proportion of correctly classified samples among all instances. ACC = (TP + TN) / (P + N) (ACC: accuracy, P: positive, N: negative) The F1 score reflects the harmonic mean of sensitivity and precision, assessing the overall performance of the model. F1 = 2TP / (2TP + FP + FN). The deep learning process was performed on a CPU with 16 GB RAM (Vostro 3520, Dell Inc., Texas, USA) for both classification and object detection tasks. Results Classification Model In this study, 812 labels on 812 images were used for training by a restorative dentistry specialist and an oral radiologist, and 204 labels on 204 images were used for testing (Table 1 ). Each cropped image belonged to either the interproximal caries class or the cervical burnout class. The model's accuracy rate (accuracy): 0.9314, sensitivity (recall): 0.8642, specificity (specificity): 0.9756, precision (precision): 0.9589, F1 score: 0.9091 (Fig. 5 ). Object Detection Model In this study, 812 labels in 364 images were used for training by a restorative dentistry specialist and an oral radiologist, and 204 labels in 90 images were used for testing (Table 1 ). The model's accuracy rate (accuracy) is 0.8174, the F1 score is 0.8487, and the mAP at the 0.5 IoU threshold was 0.82 (Fig. 6 ). Discussion This study aimed to evaluate the effectiveness of deep learning-based CNNs in the diagnosis of cervical burnout and interproximal caries on bitewing radiographs. A classification model and an object detection model were developed using 454 bitewing radiographs obtained from the archives of the Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, XXX University. It could tell the difference between interproximal caries and cervical burnout with 93.14% accuracy, 86.42% sensitivity, and 97.56% specificity. The object detection model, on the other hand, achieved 81.74% accuracy and 0.82 mAP (at 0.5 IoU threshold) performance. These findings suggest that deep learning techniques can be used as a reliable and rapid diagnostic tool in dental radiographic analysis. Deep learning stands out in dental imaging with its ability to automatically learn complex data relationships.[ 17 ] Our research shows that CNN-based models may achieve excellent accuracy in differentiating cervical burnout and interproximal caries. This presents considerable potential in clinical applications, as cervical burnout is an artifact stemming from discrepancies in X-ray absorption between radiopaque structures and is frequently misidentified as caries while not being an actual pathology. Distinguishing cervical burnout from caries can provide a significant challenge, particularly for dental students and naive clinicians. The models we developed may reduce such confusion and provide clinicians with a more consistent and objective assessment. Moreover, the capacity of deep learning algorithms to swiftly examine extensive data sets conserves time and reduces the margin of error in contrast to conventional manual techniques. [ 18 ] In the future, these models are expected to assist in treatment planning and patient follow-up activities. The logic of our classification and object detection models played an important role in the design of this study. Classification models are based on the principle that the algorithm chooses one of the presented options, and therefore, the images are cropped to include only the target region.[ 19 ] For example, in our study, only the areas where interproximal caries or cervical burnout are present, not the entire tooth, were used. This approach allows the model to focus only on the classification task; however, if both caries and cervical burnout are present on a tooth, the model may not be able to distinguish this complex situation. On the other hand, our object detection model (ResNet34) provides a more comprehensive analysis by first estimating the localization of the target region and then performing classification. Choosing the ResNet34 model, which is optimized for small data sets, allowed us to achieve satisfactory performance despite our limited data amount. This bidirectional approach shows that our study provides a strong basis in terms of both object detection and classification. There are various studies in the literature on the use of deep learning techniques in the diagnosis of dental pathologies.[ 11 , 20 , 21 ] For example, Musri et al. [ 4 ] examined the effectiveness of CNN models for the early detection of caries on periapical radiographs and reported that these models support clinical decision-making processes with accuracy rates of over 90%. Mohammad-Rahimi et al. [ 5 ] conducted a systematic review of deep learning systems for caries detection and emphasized that this technology is increasingly accepted in dental diagnosis. Putra et al. [ 7 ] discussed the general applications of artificial intelligence in digital dental radiographs and reported that deep learning exhibited superior performance, especially in areas such as caries detection, restoration evaluation, and analysis of anatomical structures. Bayrakdar et al. [ 22 ] developed a U-Net-based deep learning approach for caries detection and segmentation on bitewing radiographs and detected different caries types, including cervical caries, with a success rate of over 85% in segmentation. In addition, Bayati et al. [ 23 ] proposed a model that achieved nearly 90% accuracy in the detection of caries and other dental anomalies on panoramic radiographs. However, these studies generally focused on dental caries detection or widespread pathologies and did not focus on the separation of a specific artifact such as cervical burnout. In this context, our study fills an important gap in the literature and addresses the potential of deep learning models for the first time in the separation of cervical burnout from interproximal caries. The models we used in our study showed successful results in the detection of cervical burnout as well as interproximal caries. The uncertainty created by cervical burnout in radiographic images made it necessary for us to address a complex problem that requires both detection and classification in our study. The main limitation of our study is that our dataset includes a limited number of radiographs (454 bitewing radiographs). This may limit the generalizability of our models and their performance in different clinical scenarios, as a larger dataset could have allowed the model to learn a wider variety of pathological conditions and image variations. Also, our dataset only comes from one center and only includes patients from that center, as well as the device's specs (KaVo FOCUS, 60 kV, 7 mA, 0.08 s), and the way it was collected. Images obtained from patient groups with different radiography devices, acquisition parameters, or demographic characteristics may affect the performance of the model, which poses a weakness in terms of generalizability. Although the labeling of radiographs was performed jointly by two expert dentists, subjective interpretation differences due to human factors could not be eliminated, which could affect the consistency, especially in the identification of subtle artifacts such as cervical burnout. Also, secondary caries, caries with a lot of occlusal loss, and images with a lot of anatomical overlap were not included in our study. This choice let the model be trained with a cleaner and more standard dataset, but it meant that its performance couldn't be tested in the kind of complicated cases that happen in real clinical settings. The object detection model's lower accuracy rate (81.74%) compared to the classification model is mostly due to mistakes in localization. This suggests that the ResNet34 model may not have been able to provide enough depth in complex localization tasks, even though it was designed to work best with small datasets. In addition, technical limitations such as hardware (16 GB RAM, CPU) and training time used in our study prevented testing of more complex models (e.g., deeper variants such as ResNet50 or EfficientNet), which constituted a limitation in terms of optimizing performance. These limitations may be overcome in the future with research using larger data sets, multicenter studies, and advanced computational resources. Conclusion In conclusion, this study demonstrates that deep learning-based CNN models are a promising tool in the diagnosis of cervical burnout and interproximal caries in bitewing radiographs. Despite the limited dataset, the high accuracy rates obtained indicate that this technology can support clinicians in dental diagnostic processes. Future studies with larger and more heterogeneous datasets may increase the generalizability of the models and encourage their widespread use in clinical practice. Declarations Ethics approval and consent to participate: The study was approved by the Nuh Naci Yazgan University Non-Interventional Clinical Research Ethics Committee (decision date and number: 2025/03-001), and the study protocol was carried out following the ethical guidelines stated in the Declaration of Helsinki. Informed consent forms were obtained from the patients whose radiographs were used and submitted to the ethics committee. Funding The study has been self-funded. Research Support This research received no external financial or non-financial support Authorship Contribution Statement Ozcan Karatas : Writing – review & editing, Supervision. Ridvan Akyol : Methodology, Investigation. Kemal Selcuk Yucel : Data curation, Conceptualization, Methodology. Ebru Delikan : Writing – review & editing. Ayse Tugba Erturk Avunduk : Conceptualization, Formal analysis. Declaration of Competing Interest The authors assert that they possess no declared conflicting financial interests or personal affiliations that might have seemingly affected the research presented in this study. Acknowledgments The authors declare no acknowledgments. Data availability Data availability and the datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request. References Mohammad-Rahimi H, Nadimi M, Rohban MH, Shamsoddin E, Lee VY, Motamedian SR: Machine learning and orthodontics, current trends and the future opportunities: A scoping review. Am J Orthod Dentofacial Orthop 2021, 160(2):170-192 e174. Litjens G, Kooi T, Bejnordi BE, Setio AAA, Ciompi F, Ghafoorian M, van der Laak J, van Ginneken B, Sanchez CI: A survey on deep learning in medical image analysis. Med Image Anal 2017, 42:60-88. Schmidhuber J: Deep learning in neural networks: an overview. Neural Netw 2015, 61:85-117. Musri N, Christie B, Ichwan SJA, Cahyanto A: Deep learning convolutional neural network algorithms for the early detection and diagnosis of dental caries on periapical radiographs: A systematic review. Imaging Sci Dent 2021, 51(3):237-242. Mohammad-Rahimi H, Motamedian SR, Rohban MH, Krois J, Uribe SE, Mahmoudinia E, Rokhshad R, Nadimi M, Schwendicke F: Deep learning for caries detection: A systematic review. J Dent 2022, 122:104115. Caffery LJ, Smith AC, Bradford NK: Author's response to comment on "Accuracy of dental images for the diagnosis of dental caries and enamel defects in children and adolescents: A systematic review". J Telemed Telecare 2017, 23(5):563. Putra RH, Doi C, Yoda N, Astuti ER, Sasaki K: Current applications and development of artificial intelligence for digital dental radiography. Dentomaxillofac Radiol 2022, 51(1):20210197. Akarslan ZZ, Akdevelioglu M, Gungor K, Erten H: A comparison of the diagnostic accuracy of bitewing, periapical, unfiltered and filtered digital panoramic images for approximal caries detection in posterior teeth. Dentomaxillofac Radiol 2008, 37(8):458-463. Berry HM, Jr.: Cervical burnout and Mach band: two shadows of doubt in radiologic interpretation of carious lesions. J Am Dent Assoc 1983, 106(5):622-625. Rahmatulla M, Wyne AH: Classification of cervical burnout and its distribution in the dentition. Indian J Dent Res 1995, 6(1):13-19. Lee S, Oh SI, Jo J, Kang S, Shin Y, Park JW: Deep learning for early dental caries detection in bitewing radiographs. Sci Rep 2021, 11(1):16807. Szabo V, Szabo BT, Orhan K, Veres DS, Manulis D, Ezhov M, Sanders A: Validation of artificial intelligence application for dental caries diagnosis on intraoral bitewing and periapical radiographs. J Dent 2024, 147:105105. Chaves ET, Vinayahalingam S, van Nistelrooij N, Xi T, Romero VHD, Flügge T, Saker H, Kim A, Lima GD, Loomans B et al : Detection of caries around restorations on bitewings using deep learning. J Dent 2024, 143. Tzutalin. LabelImg. Git code (2015). https://github.com/tzutalin/labelImg. arXiv:1512.03385 [cs.CV],(or arXiv:1512.03385v1 [cs.CV] for this version), https://doi.org/10.48550/arXiv.1512.03385. Hu YF, Tang HJ, Pan G: Spiking Deep Residual Networks. Ieee T Neur Net Lear 2023, 34(8):5200-5205. Lee S, Kim D, Jeong HG: Detecting 17 fine-grained dental anomalies from panoramic dental radiography using artificial intelligence. Sci Rep 2022, 12(1):5172. Shi J, Zhang K, Guo C, Yang Y, Xu Y, Wu J: A survey of label-noise deep learning for medical image analysis. Med Image Anal 2024, 95:103166. Abdalla-Aslan R, Yeshua T, Kabla D, Leichter I, Nadler C: An artificial intelligence system using machine-learning for automatic detection and classification of dental restorations in panoramic radiography. Oral Surg Oral Med Oral Pathol Oral Radiol 2020, 130(5):593-602. Wang H, Jin Q, Li S, Liu S, Wang M, Song Z: A comprehensive survey on deep active learning in medical image analysis. Med Image Anal 2024, 95:103201. Karatas O, Cakir NN, Ozsariyildiz SS, Kis HC, Demirbuga S, Gurgan CA: A deep learning approach to dental restoration classification from bitewing and periapical radiographs. Quintessence Int 2021, 52(7):568-574. Bayrakdar IS, Orhan K, Akarsu S, Celik O, Atasoy S, Pekince A, Yasa Y, Bilgir E, Saglam H, Aslan AF et al : Deep-learning approach for caries detection and segmentation on dental bitewing radiographs. Oral Radiol 2022, 38(4):468-479. Bayati M, Alizadeh Savareh B, Ahmadinejad H, Mosavat F: Advanced AI-driven detection of interproximal caries in bitewing radiographs using YOLOv8. Sci Rep 2025, 15(1):4641. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted 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-6693428","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":467088703,"identity":"a591115b-2e71-4f7d-ab45-d75b7cbf174d","order_by":0,"name":"Ozcan KARATAS","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAyklEQVRIiWNgGAWjYJACxgYgwQ9iJRQQr8WAQRKkL8GAFC0GB0BMYrTw8x9+9nDmnj/2xudXJ354YMAgzy92AL8WyYZj5oYbnhkkbrvxdrME0GGGM2cn4NdicLCHTfLBAYMEsxtnN4C0JBjcJqTlMA9Yi73xjLObfxCn5RhQy4YDBowb+Hu3EWeLZA+bmeSMA8aJM27wbrNIMJAg7BdQiEn2HJCz5+8/u/nmjwobeX5pAloQQAKsUoJY5WD7DpCiehSMglEwCkYSAAAclUOdRp87lwAAAABJRU5ErkJggg==","orcid":"","institution":"Nuh Naci Yazgan University","correspondingAuthor":true,"prefix":"","firstName":"Ozcan","middleName":"","lastName":"KARATAS","suffix":""},{"id":467088704,"identity":"aa6a69d9-4c99-4281-abb2-877ba6338472","order_by":1,"name":"Ridvan AKYOL","email":"","orcid":"","institution":"Nuh Naci Yazgan University","correspondingAuthor":false,"prefix":"","firstName":"Ridvan","middleName":"","lastName":"AKYOL","suffix":""},{"id":467088705,"identity":"b0b4100e-cc63-4072-9e5d-3a28880ccef5","order_by":2,"name":"Kemal Selcuk YUCEL","email":"","orcid":"","institution":"Erciyes University","correspondingAuthor":false,"prefix":"","firstName":"Kemal","middleName":"Selcuk","lastName":"YUCEL","suffix":""},{"id":467088706,"identity":"f841fd58-595c-4b44-896f-a4437d5d392d","order_by":3,"name":"Ebru DELIKAN","email":"","orcid":"","institution":"Nuh Naci Yazgan University","correspondingAuthor":false,"prefix":"","firstName":"Ebru","middleName":"","lastName":"DELIKAN","suffix":""},{"id":467088707,"identity":"5e7494d2-4cc7-48c3-9025-be21f05f5178","order_by":4,"name":"Ayse Tugba Erturk AVUNDUK","email":"","orcid":"","institution":"Mersin University","correspondingAuthor":false,"prefix":"","firstName":"Ayse","middleName":"Tugba Erturk","lastName":"AVUNDUK","suffix":""}],"badges":[],"createdAt":"2025-05-18 19:23:16","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6693428/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6693428/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":84341219,"identity":"b5d2c466-3aec-4a63-a9dd-deb9d6d973ac","added_by":"auto","created_at":"2025-06-10 18:30:37","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1613093,"visible":true,"origin":"","legend":"\u003cp\u003eLabeling of interproximal caries and cervical burnout images for the object detection task with LabelImg software. The process of labeling interproximal caries separately as rectangular ROIs on the same bitewing image is shown in A, B, C, D, E, and F. The process of labeling cervical burnout images as rectangular ROIs on G. The process of labeling both interproximal caries and cervical burnout images separately as rectangular ROIs on the same bitewing image is shown in H and I.\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-6693428/v1/8cec8735725b9d4829b386a8.png"},{"id":84339371,"identity":"f80e36dc-23cc-40a5-9b35-125feee0bd5b","added_by":"auto","created_at":"2025-06-10 18:14:37","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":312186,"visible":true,"origin":"","legend":"\u003cp\u003eExporting labeled interproximal caries and cervical burnout ROIs for the classification task. The exporting process of interproximal caries is shown in A, B, C, and D. The Exporting process of cervical burnout images is shown in E, F, G, and H.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-6693428/v1/b5e0fe368e7516d790a85857.png"},{"id":84339382,"identity":"74b690eb-62f3-4e8a-96d6-931b54522f77","added_by":"auto","created_at":"2025-06-10 18:14:37","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":337953,"visible":true,"origin":"","legend":"\u003cp\u003eCNN Model created for the classification task.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-6693428/v1/7c82c26dacc46fdca6b28f46.png"},{"id":84339378,"identity":"b928da25-51a8-47f9-a590-f20228470198","added_by":"auto","created_at":"2025-06-10 18:14:37","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":432367,"visible":true,"origin":"","legend":"\u003cp\u003eResNet34 Model created for the Object Detection task.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-6693428/v1/c6a02f5e74a9ac9b5859d3b6.png"},{"id":84339377,"identity":"f09b8ea6-8b68-450d-8a32-101f1da9df67","added_by":"auto","created_at":"2025-06-10 18:14:37","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":395007,"visible":true,"origin":"","legend":"\u003cp\u003eRepresentative images of correctly predicted images from the CNN Model. 0= Interproximal Caries, 1= Cervical Burnout.\u003c/p\u003e","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-6693428/v1/b1ca0ffcfc93091be32d3594.png"},{"id":84340013,"identity":"575fa0dc-9b0b-4dfa-9ae8-f232cf9964cf","added_by":"auto","created_at":"2025-06-10 18:22:37","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":1620768,"visible":true,"origin":"","legend":"\u003cp\u003eRepresentative images of ResNet34 Model Predictions. The detection predictions shown in A, B, and C match all the labels in the labeled images. Although the detection shown in D was correctly detected as cervical burnout, the cervical burnout image shown with the blue arrow could not be detected. In the detection shown in E, incorrect localization was detected and the interproximal caries in the image labeled with the green arrow could not be detected. In the detection shown in F, while the adjacent enamel caries were detected, the interproximal caries in the labeled image shown with the yellow arrow could not be detected. The detection shown in G was detected as cervical burnout, although it was labeled as interproximal caries.\u003c/p\u003e","description":"","filename":"Figure6.png","url":"https://assets-eu.researchsquare.com/files/rs-6693428/v1/578a742b19847194c2d978a9.png"},{"id":87700031,"identity":"e71e1670-c353-4bcf-b75e-75ebabdc8f41","added_by":"auto","created_at":"2025-07-28 07:09:57","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":5724608,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6693428/v1/7453d7a5-e088-4163-afa7-69c5ec5bea6e.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003eDeep Learning-based Diagnosis of Cervical Burnout and Interproximal Caries in Bitewing Radiographs\u003c/p\u003e","fulltext":[{"header":"Background","content":"\u003cp\u003eAdvancements in artificial intelligence (AI) applications are currently resulting in significant changes in the field of health sciences. [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e] Machine learning (ML) is a highly dynamic and swiftly evolving branch of artificial intelligence predicated on the concept of empowering computers to learn from data. Deep learning, a subset of machine learning, employs multi-layered artificial neural networks to discern intricate data relationships. Deep neural networks, modeled after the functionality of human brain neurons, process data hierarchically to extract advanced features. Unlike conventional image processing and machine learning methods that necessitate manual feature extraction, deep learning models, particularly Convolutional Neural Networks (CNN), autonomously learn these features, enhancing the efficiency of the analysis process. [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e] Therefore, CNNs are extensively utilized in dentistry owing to their exceptional efficacy, particularly in image-based evaluations. Research indicates that deep learning-based CNNs yield effective outcomes in various domains, including tooth classification from 2D and 3D images, detection of caries and lesions, assessment of anatomical structures, orthodontic evaluations, and identification of restorations. [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e] Traditional manual evaluation methods can be challenging to use and take a long time. Deep learning systems, on the other hand, can quickly and effectively analyze large datasets to help clinicians make decisions. Furthermore, the implementation of such systems enhances the reliability and consistency of diagnoses for patients by minimizing the margin of error in dentistry. [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]\u003c/p\u003e \u003cp\u003eDental caries are among the most prevalent ailments encountered in society. Early diagnosis and treatment of caries is fundamental to protecting oral and dental health. [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] Alongside intraoral, visual, and tactile examinations, both traditional and digital radiography are commonly used in the identification of dental caries. Currently, digital radiographs are increasingly favored over traditional radiography owing to numerous advantages, including reduced application time, enhanced storage capabilities, and improved data processing and transfer. [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] Bite-wing radiographs are among the most frequently utilized diagnostic techniques, particularly for identifying interproximal caries. Radiographs play a crucial role in diagnostic dentistry as they reveal details that are invisible to the human eye, such as interproximal cavities, bone levels, existing restorations, and secondary cavities. [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e] Nonetheless, artifacts seen on radiographs might hinder clinicians' evaluations, make it challenging to evaluate the patient, and may even result in an inaccurate diagnosis. A phenomenon known as cervical burnout is a radiolucent band or wedge-shaped structure seen in the cervical areas of teeth on radiographs. [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] Cervical burnout, an artifact resulting from the differential absorption of X-rays by radiopaque structures like enamel, dentin, cementum, and bone, is not a genuine pathology. However, it is a concern highlighted that practitioners should not conflate this artifact with interproximal caries. Likewise, burnouts observed beneath interproximal restorations may be confused with secondary caries. [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]\u003c/p\u003e \u003cp\u003eTo accurately tell the difference between interproximal caries and cervical burnout, it is suggested that the clinical and radiographic evaluations be done at the same time. In such cases, assessments conducted on radiographic images utilizing supportive AI systems may help clinicians in achieving accurate diagnoses and save time. Therefore, assessing the effectiveness of AI applications on this topic will be an essential issue of discussion. Unfortunately, to the best of our knowledge, no study has been conducted in the literature to investigate the effectiveness of deep learning-based ResNets in detecting cervical burnout and their potential for differential diagnosis with caries. Within this framework, our objective is to assess the effectiveness of ResNets in the diagnosis of cervical burnout and interproximal caries in bitewing radiographs.\u003c/p\u003e \u003cp\u003eSince cervical burnout and interproximal caries have similar appearances, problems may occur both in the object detection phase and in the classification phase, and the detected lesion may be classified incorrectly. Especially in the studies of Lee et al.[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] using the Inceptionv3 model, the False Negative rates primarily led us to the classification task. Therefore, we examined the classification success with the CNN model, which has proven its performance in classification tasks in dental radiographs.[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] Although there are studies in the literature with CNN-based models such as Mask R-CNN with small data sets, we did not come across a study on the detection algorithm with the ResNet model prepared for small data sets. [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003eStudy design and ethical approval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis retrospective study was conducted using a dataset consisting of bitewing radiographs available in the archives of the Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, Nuh Naci Yazgan University. The study was approved by the Nuh Naci Yazgan University Non-Interventional Clinical Research Ethics Committee (decision date and number: 2025/03-001), and the study protocol was carried out following the ethical guidelines stated in the Declaration of Helsinki.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDataset Creation and Image Preprocessing\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eRadiographs were randomly selected by two dentists from the archives of the Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, XXX University, containing bitewing radiographs taken between January 2024 and January 2025 for the diagnosis and treatment of caries. Only radiographs containing permanent teeth were used as data, without any additional information about the patients (e.g., gender, age, or other clinical information). Radiographs with poor image quality, excessive noise, artifacts, or severe overlapping of proximal surfaces due to anatomical alignment of certain teeth were excluded. In addition, all radiographic images used in this study were obtained using the same device (KaVo X-ray machine; KaVo FOCUS, Tuusula,\u0026nbsp;Finland) using the irradiation parameters of 60 kV, 7 mA, and 0.08 s.\u003c/p\u003e\n\u003cp\u003eA restorative dentist with at least 10 years of experience and an oral radiologist were able to label the radiograph images together, and images that could not be agreed upon were excluded. Since the aim was to distinguish between interproximal caries and cervical burnouts on bitewing images, it was decided to create both a classification model (Classification) and an object detection model (Detection). Then, only interproximal caries and cervical burnout images were labeled for the object detection model, and LabelImg software was used for labeling (Copyright (C) 2013 MIT, Computer Science and Artificial Intelligence Laboratory) [14] (Figure 1). Secondary caries under restoration and caries with high occlusal loss were excluded from the study to not affect performance. At the same time, the labeled regions of interest (ROI) were cropped and exported for the classification task (Figure 2). All bitewing images were exported in PNG format, while the labels were saved in XML (Extensible Markup Language) format. The final dataset consisted of 454 bitewing radiographs, 614 interproximal caries, and 402 cervical burnout images.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eNetwork Training\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eBefore training the classification model, preprocessing steps were performed, which included resizing the images to 300 \u0026times; 300 pixels. The dataset was then split into two groups; for interproximal caries, the training set consisted of 491 images (80%) and the test set consisted of 123 images (20%). For cervical burnouts, the training set consisted of 321 images (80%), and the test set consisted of 81 images (20%).\u003c/p\u003e\n\u003cp\u003eBefore training the object detection model, preprocessing steps were performed, which included resizing the images to 600 \u0026times; 600 pixels. The dataset was then split into three separate groups, the training set consisting of 364 images (80%) and the test set consisting of 90 images (20%) (Table 1).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eTable1.\u003c/strong\u003e Dataset and label distribution\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" align=\"\" class=\"fr-table-selection-hover\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDataset Total Bitewing Radiographs\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 377px;\"\u003e\n \u003cp\u003e454\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTotal Interproximal Caries\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 377px;\"\u003e\n \u003cp\u003e614\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTotal Cervical Burnout\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 377px;\"\u003e\n \u003cp\u003e402\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTotal Number of Labels\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 377px;\"\u003e\n \u003cp\u003e1016\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e\u003cu\u003eTraining Set\u003c/u\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e\u003cu\u003eTest Set\u003c/u\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eClassification Model\u003cbr\u003e\u0026nbsp;CNN/ Number of Labels\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e491 + 321\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e123 + 81\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eClassification Model\u003cbr\u003e\u0026nbsp;CNN/ Number of Images\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e491 + 321\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e123 + 81\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eObject Detection Task\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eResNet34/ Number of Labels\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e491 + 321\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e123 + 81\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eObject Detection Task ResNet34\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e364\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 188px;\"\u003e\n \u003cp\u003e90\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eDeep Learning Procedure\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFor the classification task, a 32-layer CNN model was created with the Keras application from the TensorFlow library. While the activation function \u0026ldquo;SoftMax\u0026rdquo; was used in the last layer, \u0026ldquo;ReLU\u0026rdquo; was used in the other layers. The parameters \u0026ldquo;binary_crossentropy,\u0026rdquo; \u0026ldquo;rmsprop,\u0026rdquo; and 32 as batch_size were used as the loss function. The evaluation metrics were carried out using the confusion matrix (Figure 3).\u003c/p\u003e\n\u003cp\u003eA ResNet34 model (Residual Network) was created for the object detection task via PyTorch (Figure 4). With the torchvision.models.resnet34 function, the weights parameter ResNet34_Weights is optional; the process parameter (progress) is a bool, optional, and the kwargs parameters are any; the optimizer parameter is \u0026quot;Adam,\u0026quot; the learning rate parameter is \u0026quot;0.001,\u0026quot; and the batch size parameter was determined as 64. [15]\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResNet34 model (Residual Network)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIntroduced by Microsoft Research in 2015, ResNet brought a revolutionary approach to deep learning by utilizing residual blocks to facilitate the training of deeper networks. It is a state-of-the-art CNN architecture that enables networks to reach substantial depths while achieving remarkable success in numerous applications. Residual connections are employed to address challenges encountered during the training of deep neural networks, such as the vanishing gradient problem. This approach allows for the construction of more powerful and complex models by adding additional layers. [16]\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMetrics for the model\u0026rsquo;s performance\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe evaluation metrics were determined as sensitivity, specificity, precision, accuracy, and F1 score.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eSensitivity (TPR)\u003c/em\u003e\u003c/strong\u003e represents the true positive rate and indicates the likelihood of correctly identifying a disease. \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTPR = TP / (TP + FN) (TPR: true positive rate, TP: true positive, FN: false negative)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eSpecificity (SPC)\u003c/em\u003e\u003c/strong\u003e represents the true negative rate and indicates the likelihood of correctly identifying healthy individuals.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eSPC = TN / (FP + TN) (SPC: specificity, TN: true negative, FP: false positive)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003ePrecision (PPV)\u003c/em\u003e\u003c/strong\u003e is the positive predictive value and shows the probability that identified positives are truly diseased.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003ePPV = TP / (TP + FP) (PPV: positive predictive value)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e\u0026nbsp;\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Accuracy (ACC)\u003c/em\u003e\u003c/strong\u003e represents the proportion of correctly classified samples among all instances.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;ACC = (TP + TN) / (P + N) (ACC: accuracy, P: positive, N: negative)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eThe F1 score\u003c/em\u003e\u003c/strong\u003e reflects the harmonic mean of sensitivity and precision, assessing the overall performance of the model.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eF1 = 2TP / (2TP + FP + FN).\u003c/p\u003e\n\u003cp\u003eThe deep learning process was performed on a CPU with 16 GB RAM (Vostro 3520, Dell Inc., Texas, USA) for both classification and object detection tasks.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003eClassification Model\u003c/h2\u003e \u003cp\u003eIn this study, 812 labels on 812 images were used for training by a restorative dentistry specialist and an oral radiologist, and 204 labels on 204 images were used for testing (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Each cropped image belonged to either the interproximal caries class or the cervical burnout class. The model's accuracy rate (accuracy): 0.9314, sensitivity (recall): 0.8642, specificity (specificity): 0.9756, precision (precision): 0.9589, F1 score: 0.9091 (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eObject Detection Model\u003c/h2\u003e \u003cp\u003eIn this study, 812 labels in 364 images were used for training by a restorative dentistry specialist and an oral radiologist, and 204 labels in 90 images were used for testing (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). The model's accuracy rate (accuracy) is 0.8174, the F1 score is 0.8487, and the mAP at the 0.5 IoU threshold was 0.82 (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study aimed to evaluate the effectiveness of deep learning-based CNNs in the diagnosis of cervical burnout and interproximal caries on bitewing radiographs. A classification model and an object detection model were developed using 454 bitewing radiographs obtained from the archives of the Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, XXX University. It could tell the difference between interproximal caries and cervical burnout with 93.14% accuracy, 86.42% sensitivity, and 97.56% specificity. The object detection model, on the other hand, achieved 81.74% accuracy and 0.82 mAP (at 0.5 IoU threshold) performance. These findings suggest that deep learning techniques can be used as a reliable and rapid diagnostic tool in dental radiographic analysis.\u003c/p\u003e \u003cp\u003eDeep learning stands out in dental imaging with its ability to automatically learn complex data relationships.[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] Our research shows that CNN-based models may achieve excellent accuracy in differentiating cervical burnout and interproximal caries. This presents considerable potential in clinical applications, as cervical burnout is an artifact stemming from discrepancies in X-ray absorption between radiopaque structures and is frequently misidentified as caries while not being an actual pathology. Distinguishing cervical burnout from caries can provide a significant challenge, particularly for dental students and naive clinicians. The models we developed may reduce such confusion and provide clinicians with a more consistent and objective assessment. Moreover, the capacity of deep learning algorithms to swiftly examine extensive data sets conserves time and reduces the margin of error in contrast to conventional manual techniques. [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] In the future, these models are expected to assist in treatment planning and patient follow-up activities.\u003c/p\u003e \u003cp\u003eThe logic of our classification and object detection models played an important role in the design of this study. Classification models are based on the principle that the algorithm chooses one of the presented options, and therefore, the images are cropped to include only the target region.[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] For example, in our study, only the areas where interproximal caries or cervical burnout are present, not the entire tooth, were used. This approach allows the model to focus only on the classification task; however, if both caries and cervical burnout are present on a tooth, the model may not be able to distinguish this complex situation. On the other hand, our object detection model (ResNet34) provides a more comprehensive analysis by first estimating the localization of the target region and then performing classification. Choosing the ResNet34 model, which is optimized for small data sets, allowed us to achieve satisfactory performance despite our limited data amount. This bidirectional approach shows that our study provides a strong basis in terms of both object detection and classification.\u003c/p\u003e \u003cp\u003eThere are various studies in the literature on the use of deep learning techniques in the diagnosis of dental pathologies.[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] For example, Musri et al. [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e] examined the effectiveness of CNN models for the early detection of caries on periapical radiographs and reported that these models support clinical decision-making processes with accuracy rates of over 90%. Mohammad-Rahimi et al. [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e] conducted a systematic review of deep learning systems for caries detection and emphasized that this technology is increasingly accepted in dental diagnosis. Putra et al. [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] discussed the general applications of artificial intelligence in digital dental radiographs and reported that deep learning exhibited superior performance, especially in areas such as caries detection, restoration evaluation, and analysis of anatomical structures. Bayrakdar et al. [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] developed a U-Net-based deep learning approach for caries detection and segmentation on bitewing radiographs and detected different caries types, including cervical caries, with a success rate of over 85% in segmentation. In addition, Bayati et al. [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] proposed a model that achieved nearly 90% accuracy in the detection of caries and other dental anomalies on panoramic radiographs. However, these studies generally focused on dental caries detection or widespread pathologies and did not focus on the separation of a specific artifact such as cervical burnout. In this context, our study fills an important gap in the literature and addresses the potential of deep learning models for the first time in the separation of cervical burnout from interproximal caries. The models we used in our study showed successful results in the detection of cervical burnout as well as interproximal caries. The uncertainty created by cervical burnout in radiographic images made it necessary for us to address a complex problem that requires both detection and classification in our study.\u003c/p\u003e \u003cp\u003eThe main limitation of our study is that our dataset includes a limited number of radiographs (454 bitewing radiographs). This may limit the generalizability of our models and their performance in different clinical scenarios, as a larger dataset could have allowed the model to learn a wider variety of pathological conditions and image variations. Also, our dataset only comes from one center and only includes patients from that center, as well as the device's specs (KaVo FOCUS, 60 kV, 7 mA, 0.08 s), and the way it was collected. Images obtained from patient groups with different radiography devices, acquisition parameters, or demographic characteristics may affect the performance of the model, which poses a weakness in terms of generalizability. Although the labeling of radiographs was performed jointly by two expert dentists, subjective interpretation differences due to human factors could not be eliminated, which could affect the consistency, especially in the identification of subtle artifacts such as cervical burnout. Also, secondary caries, caries with a lot of occlusal loss, and images with a lot of anatomical overlap were not included in our study. This choice let the model be trained with a cleaner and more standard dataset, but it meant that its performance couldn't be tested in the kind of complicated cases that happen in real clinical settings. The object detection model's lower accuracy rate (81.74%) compared to the classification model is mostly due to mistakes in localization. This suggests that the ResNet34 model may not have been able to provide enough depth in complex localization tasks, even though it was designed to work best with small datasets. In addition, technical limitations such as hardware (16 GB RAM, CPU) and training time used in our study prevented testing of more complex models (e.g., deeper variants such as ResNet50 or EfficientNet), which constituted a limitation in terms of optimizing performance. These limitations may be overcome in the future with research using larger data sets, multicenter studies, and advanced computational resources.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn conclusion, this study demonstrates that deep learning-based CNN models are a promising tool in the diagnosis of cervical burnout and interproximal caries in bitewing radiographs. Despite the limited dataset, the high accuracy rates obtained indicate that this technology can support clinicians in dental diagnostic processes. Future studies with larger and more heterogeneous datasets may increase the generalizability of the models and encourage their widespread use in clinical practice.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate:\u003c/strong\u003e The study was approved by the Nuh Naci Yazgan University Non-Interventional Clinical Research Ethics Committee (decision date and number: 2025/03-001), and the study protocol was carried out following the ethical guidelines stated in the Declaration of Helsinki. Informed consent forms were obtained from the patients whose radiographs were used and submitted to the ethics committee.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study has been self-funded.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResearch Support\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research received no external financial or non-financial support\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthorship Contribution Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eOzcan Karatas\u003c/strong\u003e: Writing \u0026ndash; review \u0026amp; editing, Supervision. \u003cstrong\u003eRidvan Akyol\u003c/strong\u003e: Methodology, Investigation. \u003cstrong\u003eKemal Selcuk Yucel\u003c/strong\u003e: Data curation, Conceptualization, Methodology. \u003cstrong\u003eEbru Delikan\u003c/strong\u003e: Writing \u0026ndash; review \u0026amp; editing. \u003cstrong\u003eAyse Tugba Erturk Avunduk\u003c/strong\u003e: Conceptualization, Formal analysis.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration of Competing Interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors assert that they possess no declared conflicting financial interests or personal affiliations that might have seemingly affected the research presented in this study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no acknowledgments. \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData availability and the datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eMohammad-Rahimi H, Nadimi M, Rohban MH, Shamsoddin E, Lee VY, Motamedian SR: Machine learning and orthodontics, current trends and the future opportunities: A scoping review. \u003cem\u003eAm J Orthod Dentofacial Orthop \u003c/em\u003e2021, 160(2):170-192 e174.\u003c/li\u003e\n\u003cli\u003eLitjens G, Kooi T, Bejnordi BE, Setio AAA, Ciompi F, Ghafoorian M, van der Laak J, van Ginneken B, Sanchez CI: A survey on deep learning in medical image analysis. \u003cem\u003eMed Image Anal \u003c/em\u003e2017, 42:60-88.\u003c/li\u003e\n\u003cli\u003eSchmidhuber J: Deep learning in neural networks: an overview. \u003cem\u003eNeural Netw \u003c/em\u003e2015, 61:85-117.\u003c/li\u003e\n\u003cli\u003eMusri N, Christie B, Ichwan SJA, Cahyanto A: Deep learning convolutional neural network algorithms for the early detection and diagnosis of dental caries on periapical radiographs: A systematic review. \u003cem\u003eImaging Sci Dent \u003c/em\u003e2021, 51(3):237-242.\u003c/li\u003e\n\u003cli\u003eMohammad-Rahimi H, Motamedian SR, Rohban MH, Krois J, Uribe SE, Mahmoudinia E, Rokhshad R, Nadimi M, Schwendicke F: Deep learning for caries detection: A systematic review. \u003cem\u003eJ Dent \u003c/em\u003e2022, 122:104115.\u003c/li\u003e\n\u003cli\u003eCaffery LJ, Smith AC, Bradford NK: Author\u0026apos;s response to comment on \u0026quot;Accuracy of dental images for the diagnosis of dental caries and enamel defects in children and adolescents: A systematic review\u0026quot;. \u003cem\u003eJ Telemed Telecare \u003c/em\u003e2017, 23(5):563.\u003c/li\u003e\n\u003cli\u003ePutra RH, Doi C, Yoda N, Astuti ER, Sasaki K: Current applications and development of artificial intelligence for digital dental radiography. \u003cem\u003eDentomaxillofac Radiol \u003c/em\u003e2022, 51(1):20210197.\u003c/li\u003e\n\u003cli\u003eAkarslan ZZ, Akdevelioglu M, Gungor K, Erten H: A comparison of the diagnostic accuracy of bitewing, periapical, unfiltered and filtered digital panoramic images for approximal caries detection in posterior teeth. \u003cem\u003eDentomaxillofac Radiol \u003c/em\u003e2008, 37(8):458-463.\u003c/li\u003e\n\u003cli\u003eBerry HM, Jr.: Cervical burnout and Mach band: two shadows of doubt in radiologic interpretation of carious lesions. \u003cem\u003eJ Am Dent Assoc \u003c/em\u003e1983, 106(5):622-625.\u003c/li\u003e\n\u003cli\u003eRahmatulla M, Wyne AH: Classification of cervical burnout and its distribution in the dentition. \u003cem\u003eIndian J Dent Res \u003c/em\u003e1995, 6(1):13-19.\u003c/li\u003e\n\u003cli\u003eLee S, Oh SI, Jo J, Kang S, Shin Y, Park JW: Deep learning for early dental caries detection in bitewing radiographs. \u003cem\u003eSci Rep \u003c/em\u003e2021, 11(1):16807.\u003c/li\u003e\n\u003cli\u003eSzabo V, Szabo BT, Orhan K, Veres DS, Manulis D, Ezhov M, Sanders A: Validation of artificial intelligence application for dental caries diagnosis on intraoral bitewing and periapical radiographs. \u003cem\u003eJ Dent \u003c/em\u003e2024, 147:105105.\u003c/li\u003e\n\u003cli\u003eChaves ET, Vinayahalingam S, van Nistelrooij N, Xi T, Romero VHD, Fl\u0026uuml;gge T, Saker H, Kim A, Lima GD, Loomans B\u003cem\u003e et al\u003c/em\u003e: Detection of caries around restorations on bitewings using deep learning. \u003cem\u003eJ Dent \u003c/em\u003e2024, 143.\u003c/li\u003e\n\u003cli\u003eTzutalin. LabelImg. Git code (2015). https://github.com/tzutalin/labelImg.\u003c/li\u003e\n\u003cli\u003earXiv:1512.03385 [cs.CV],(or arXiv:1512.03385v1 [cs.CV] for this version), https://doi.org/10.48550/arXiv.1512.03385.\u003c/li\u003e\n\u003cli\u003eHu YF, Tang HJ, Pan G: Spiking Deep Residual Networks. \u003cem\u003eIeee T Neur Net Lear \u003c/em\u003e2023, 34(8):5200-5205.\u003c/li\u003e\n\u003cli\u003eLee S, Kim D, Jeong HG: Detecting 17 fine-grained dental anomalies from panoramic dental radiography using artificial intelligence. \u003cem\u003eSci Rep \u003c/em\u003e2022, 12(1):5172.\u003c/li\u003e\n\u003cli\u003eShi J, Zhang K, Guo C, Yang Y, Xu Y, Wu J: A survey of label-noise deep learning for medical image analysis. \u003cem\u003eMed Image Anal \u003c/em\u003e2024, 95:103166.\u003c/li\u003e\n\u003cli\u003eAbdalla-Aslan R, Yeshua T, Kabla D, Leichter I, Nadler C: An artificial intelligence system using machine-learning for automatic detection and classification of dental restorations in panoramic radiography. \u003cem\u003eOral Surg Oral Med Oral Pathol Oral Radiol \u003c/em\u003e2020, 130(5):593-602.\u003c/li\u003e\n\u003cli\u003eWang H, Jin Q, Li S, Liu S, Wang M, Song Z: A comprehensive survey on deep active learning in medical image analysis. \u003cem\u003eMed Image Anal \u003c/em\u003e2024, 95:103201.\u003c/li\u003e\n\u003cli\u003eKaratas O, Cakir NN, Ozsariyildiz SS, Kis HC, Demirbuga S, Gurgan CA: A deep learning approach to dental restoration classification from bitewing and periapical radiographs. \u003cem\u003eQuintessence Int \u003c/em\u003e2021, 52(7):568-574.\u003c/li\u003e\n\u003cli\u003eBayrakdar IS, Orhan K, Akarsu S, Celik O, Atasoy S, Pekince A, Yasa Y, Bilgir E, Saglam H, Aslan AF\u003cem\u003e et al\u003c/em\u003e: Deep-learning approach for caries detection and segmentation on dental bitewing radiographs. \u003cem\u003eOral Radiol \u003c/em\u003e2022, 38(4):468-479.\u003c/li\u003e\n\u003cli\u003eBayati M, Alizadeh Savareh B, Ahmadinejad H, Mosavat F: Advanced AI-driven detection of interproximal caries in bitewing radiographs using YOLOv8. \u003cem\u003eSci Rep \u003c/em\u003e2025, 15(1):4641.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Cervical Burnout, Convolutional Neural Networks, Deep Learning, Interproximal Caries","lastPublishedDoi":"10.21203/rs.3.rs-6693428/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6693428/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground: \u003c/strong\u003eThis study aimed to evaluate the success of deep learning-based convolutional neural networks (CNN) and Residual Neural Network-34 (ResNet34) in classifying and detecting cervical burnout and interproximal caries on bitewing radiographs.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods: \u003c/strong\u003eWithin the scope of the study, a dataset consisting of 454 bitewing radiographs, free of noise and artifacts, was labeled by two dentists with LabelImg software. A 32-layer CNN model (614 interproximal caries, 402 cervical burnout) was created for classification, and a ResNet34 model was created for object detection. The images were resized to 300x300 pixels, and the datasets were divided into 80% training and 20% testing. Performance metrics included sensitivity, specificity, precision, accuracy, and F1 score.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults: \u003c/strong\u003eThe classification model achieved 93.14% accuracy, 86.42% sensitivity, and 97.56% specificity, while the object detection model gave 81.74% accuracy and 0.82 mAP values ​​at 0.5 IoU. The data showed that CNN models were successful in classifying cervical burnout and interproximal caries. The data also showed that ResNet-34 models were successful in detecting cervical burnout and interproximal caries.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion: \u003c/strong\u003eDespite a limited dataset, CNN models showed successful results in classifying cervical burnout and interproximal caries on bitewing radiographs.\u003c/p\u003e","manuscriptTitle":"Deep Learning-based Diagnosis of Cervical Burnout and Interproximal Caries in Bitewing Radiographs","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-10 18:14:32","doi":"10.21203/rs.3.rs-6693428/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"216a0494-bc81-4292-9e75-faf3194f6d2f","owner":[],"postedDate":"June 10th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-07-28T07:09:34+00:00","versionOfRecord":[],"versionCreatedAt":"2025-06-10 18:14:32","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6693428","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6693428","identity":"rs-6693428","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}