Deep Learning-Based Fractured Tooth Detection in Occlusal Radiographs

Deep Learning-Based Fractured Tooth Detection in Occlusal 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 Fractured Tooth Detection in Occlusal Radiographs Ahmed NUSARI, Esra ONCU, Emin Argun ORAL, Ibrahim Yucel OZBEK, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7207126/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 18 Dec, 2025 Read the published version in BMC Oral Health → Version 1 posted 14 You are reading this latest preprint version Abstract Background This study highlights the importance of accurate and quick diagnosis of detection and classification of fractured teeth and the potential benefits of applying deep learning (DL) techniques to solve this problem. Methods In the study, occlusal radiography (OR) of the premaxilla is used for tooth fracture detection of teeth numbers 11 and 21. For that, a dataset that contains 200 ORs of various tooth conditions was constructed. In the proposed method, teeth numbers 11 and 21 are automatically detected in OR images using a YOLOv8-based machine learning framework as the first step. Then, images of these two teeth are obtained by cropping OR images using the bounding box coordinates of numbers 11 and 21 teeth, obtained by a YOLO-based detector. Finally, these cropped images are provided as input to a pre-trained CNN-based network to classify between the “fracture” or “non-fractured tooth”. For this purpose, VGG19, EfficientNetB0, InceptionResNetV2, and InceptionV3 nets are employed, and the obtained classification results are fused by applying a majority voting step to further improve the performance. Results The experimental studies obtained a 99.5% mean average precision (mAP50) score. On the other hand, percent accuracy rates in the range from 84.65 to 87.92 were observed using five pre-trained networks, and the percent accuracy metric was improved to 91.94% using the majority voting-based fusion approach. Conclusion The findings indicate that the proposed method effectively detects fractured teeth by leveraging machine learning techniques. Furthermore, this approach provides a pioneering framework for integrating artificial intelligence (AI) methodologies into dental diagnostics, offering clinicians a reliable decision-support tool for improved diagnostic accuracy. Dental Tooth Fracture Tooth Detection CNN models YOLOv8 Occlusal Radiography Majority Voting Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Introduction Traumatic dental injuries constitute a significant concern in dentistry, as they can lead to functional, aesthetic, and psychological complications. These injuries are particularly prevalent in maxillary central incisors, which hold critical importance in both functional and aesthetic aspects [ 1 ]. Dental trauma may result in crown or root fractures independently, or in some cases, both types of fractures may occur simultaneously. Early diagnosis and appropriate management through suitable treatment modalities are crucial factors in determining the prognosis of the affected tooth [ 2 ]. Radiographic evaluation should be conducted alongside clinical examination in the diagnosis of dental trauma, with careful assessment of the fracture's size, location, and the extent of damage to the supporting tissues. Periapical, panoramic, and occlusal radiographs (OR) are commonly used imaging modalities in clinical practice for diagnosing dental fractures [ 3 ]. In addition to these, cone-beam computed tomography (CBCT) is an advanced imaging technique that provides three-dimensional evaluation, particularly in complex fracture cases. OR serves as a crucial diagnostic tool in the assessment of dental trauma and is particularly preferred in cases where periapical radiographs are insufficient or when imaging from additional angles is required [ 4 ]. The International Association of Dental Traumatology recommends obtaining at least two periapical radiographs and one OR for the evaluation of dental trauma cases, along with the examination of radiographic images taken at different vertical and horizontal angulations [ 1 ]. In recent years, artificial intelligence (AI)-powered technologies have made significant advancements in the field of medical image analysis and have increasingly been integrated into dental practice [ 5 ]. Deep learning (DL), a subcategory of AI, refers to sophisticated neural networks composed of multiple layers. Convolutional neural networks (CNNs) represent a specific type of neural network. Due to the capability of CNNs to diagnose and analyze medical images, researchers widely employ them in contemporary studies. This capability stems from the convolutional process, which involves extracting a substantial amount of pixel-level information from images, enabling the model to learn and recognize image patterns effectively [ 6 ]. Two fundamental factors play a crucial role in the interpretation of radiological images: the inclusion of diagnostic information within the image and the ability of the evaluator to accurately perceive this information. However, limitations in clinical experience among dentists or a lack of familiarity with previously unencountered cases may lead to incorrect or incomplete diagnoses [ 7 ]. At this point, the integration of AI-based systems holds the potential to support the diagnostic process by reducing clinician error rates and shortening the time required for diagnosis. This study aims to detect crown fractures in maxillary central incisors using an AI-assisted system based on OR. The proposed method consists of two stages: In the first stage, teeth are detected using the YOLOv8 model. In the second stage, the presence of crown fractures is classified using five different CNN-based transfer learning models. Following model evaluations, the majority voting method is applied to enhance diagnostic accuracy. This technique leverages a collective decision-making mechanism among multiple models, minimizing errors while improving overall accuracy. According to the literature review, this study is one of the first investigations on AI-assisted crown fracture detection using OR. Therefore, the findings have the potential to serve as a significant decision-support mechanism for clinicians in both emergency and routine dental practice. Materials and methods Data This study protocol was approved by the Ethics Committee of Atatürk University Faculty of Dentistry, Department of Oral and Maxillofacial Radiology, Erzurum (Approval No: 2023:61). OR images from 200 patients who visited between 2020 and 2022, routine clinical examination, or follow-up between 2020 and 2022 were analyzed and recorded in PNG format by a clinician. All OR images were obtained using the bisecting angle technique with a Belmont PHOT-xIIs Intraoral X-Ray unit (Takara Belmont, Somerset, MA, USA) and recorded on Size 4 (57×76 mm) phosphor plates (Dürr Dental, Bietigheim-Bissingen, Germany). The acquired images were scanned and digitized using the Vistascan Mini Easy System (Dürr Dental, Bietigheim-Bissingen, Germany). The study included cases with crown fractures in maxillary central incisors (11 and 21 according to the FDI numbering system). However, images that hindered diagnosis due to superimposition or artifacts, teeth with caries, microcrack or dental anomalies, patients with orthodontic appliances were excluded from the study. The teeth were classified as either “fractured” or “non-fractured.” Due to the avulsion of two teeth, a total of 398 teeth were evaluated: 258 were classified as fractured and 140 as non-fractured (Fig. 1 ). Figure 1 shows several images from our dataset that contain various fractured teeth. Fractured areas are denoted by orange arrows. Data annotation The labeling process was conducted using LabelImg, a software commonly utilized in DL applications. Each tooth was manually annotated on OR images and classified as either "fractured" or "non-fractured." The primary objective of this labeling process was to create a high-quality dataset for training the YOLOv8 network. The annotated images were used to enable the model to automatically detect maxillary central incisors (11 and 21). Additionally, the labeled data facilitated the automatic detection of teeth from the original radiographic images and optimized their presentation for the classification phase. Figure 2 illustrates the labeling process and how the annotated images were integrated into the AI training workflow. The proposed method For providing optimal patient care, the accurate identification of fractured teeth is essential. A more efficient and objective technique is required due to the growing need for effective dental diagnostics, as manual categorization frequently results in variances in diagnosis. In this proposed method, our primary objective could be obtaining high accuracy in the automated detection and classification of different types of tooth fractures, reducing reliance on humans in diagnosing and classifying teeth, and providing a highly accurate, repeatable, and rapid system for consistent fracture assessments. In the following sections, we will introduce the details of our method by providing its general architecture. Two networks are used, and every network has a different purpose. The first network is for detecting teeth, and the other one is for classifying whether they are fractured or not. Different image processing techniques will be applied for each network. Figure 3 illustrates the general structure of the steps for detecting and classifying teeth. A tooth detection and classification networks are trained separately. Both networks are combined and used together later. In our approach to dental image analysis, our proposed method unfolds across four integral blocks: Contrast Limited Adaptive Histogram Equalization (CLAHE) / Cropping, detection, classification, and majority voting blocks. Each contributes to a refined and accurate identification of teeth. This contribution gives us reliable results using several intelligence models. By combining these two networks, we can accurately detect and classify teeth, allowing practical analysis and diagnosis of dental conditions. CLAHE / Cropping block The initial block encompasses two image processing techniques. The first technique, CLAHE, prepares the input image for YOLOv8 training. The CLAHE method improves image quality and enhances color contrast in images by adjusting the relative resolution of lighting in image areas. Thus, the quality of the images can be improved, and the contours of the teeth can be better highlighted, which facilitates the process of accurately and effectively detecting the location of the teeth [ 8 ]. To improve the images and highlight the teeth’s contour, we applied the CLAHE method twice with varying parameters. The same CLAHE parameters are used for those two times, which are three contrast limits (clipLimit) and (45, 45) of tiles (tileGridSize) as shown in Fig. 4 . Cropping the tooth from the original image is the second method. Following YOLOv8’s tooth detection, the detected teeth’s coordinates are obtained, and a bounding box is added to the original image to point to the detected tooth. The tooth is then cropped according to the box that is painted. As a result of the cropping process, 258 fractured teeth and 140 non-fractured teeth were obtained. Due to size differences, cropped teeth were adjusted to size (448, 224, 3) to guarantee that they could be trained using CNN models as shown in Fig. 5 . This resizing step is essential to maximize the dental classification network training. Detection Block In the second block, we used YOLOv8 to provide a foundational understanding of the tooth landscape within the images and discern the distinct features of the 11 and 21 teeth. Based on this detection process, a methodical process was designed to organize the acquisition of precise tooth coordinates precisely determined by the results of YOLOv8. With the help of these obtained coordinates, individual teeth are smoothly repositioned on the original images, and cropping is performed. Ultralytics released YOLOv8 in January 2023. YOLOv8 processes regression, objectness, and classification tasks independently using an anchor-free model with a decoupled head [ 9 ]. YOLOv8 includes enhancements in the form of a new neural network architecture. This architecture contains two neural networks, the Feature Pyramid Network (FPN) and the Path Aggregation Network (PAN), built on a novel labeling tool that streamlines the annotation process. Several useful features of this labeling tool are adjustable hotkeys, automated labeling, and shortcut labeling. These characteristics together make it simple to annotate images for model training [ 10 ]. The following image processing was applied to the YOLOv8 network. To detect and locate teeth, the images used in the dataset require resizing because they have different dimensions. To use these images, they have been resized to 640 x 640 so that they are suitable for this process. Classification Block As we progress to the third block, our focus turns to harnessing CNN networks’ capabilities. These specialized models extract intricate features and classify teeth based on learned patterns. In this process, different transfer learning is performed using pre-trained CNN models. The transfer learning pre-trained CNN was tuned for one task, and its knowledge was transferred to multiple models [ 11 ]. Transfer learning is a successful approach for increasing neural network models’ capacity to detect and apply previously learned information and abilities to new models [ 12 ]. It is generally understood that optimizing and thoroughly training a model is a tough and time-consuming process that necessitates using a powerful GPU and millions of training instances. Learning to move to utilize pre-trained weights from various objective circumstances can be a useful technique for overcoming these issues [ 13 ]. VGG-19 has 19 layers, 16 convolutional layers, 3 Fully Connected layers, and five pooling layers. Batch normalization is utilized after each convolutional layer to alleviate the internal covariate shift problem and make the network quicker and more stable. Except for the last Fully Connected (FC) layer, all layers depend on the Rectified Linear Unit (ReLU) activation function. The filters are always 3x3 in size with a stride of 1. The image’s spatial resolution is preserved by using padding of size 1. Max pooling is done throughout a 2x2 window with a stride size of 2 [ 14 ]. There are 28 layers in the EfficientNetB0: 10 pooling layers, five fully connected layers, and 13 convolutional layers. The convolutional layers use A depthwise separable convolution, which lowers the amount of parameters and processing needed. Max pooling is a straightforward and effective procedure used by the pooling layers. The output from each branch is concatenated, routed through two dense layers, and then passed via a SoftMax activation function for classification. The model is trained using the Adam optimizer and the binary cross-entropy loss function [ 15 ]. InceptionResNetV2 includes 164 layers and an architecture containing global average pooling 2d, a dense layer, batch normalization, a dropout layer, an activation function, and a second dense layer to categorize the image. InceptionResNetV2 combines feature extraction from Inception with residual connections from ResNet to improve network convergence and gradient flow [ 16 ]. Convolution filters with various kernel sizes (1x1, 3x3, 5x5, and 7x7) and max pooling are used in the module. Concatenating the data enables multi-level feature extraction, which improves network performance. Later, the Inception module was merged with residual connections to generate IR v2, resulting in a more efficient and accurate model [ 17 ]. Inception V3 is a CNN-based classification network. It makes use of inception modules, which have 42 layers deep and are made up of a stacked layer with 1x1, 3 x 3, and 5x5 convolutions. Reduce the number of parameters while increasing the rate of training. This model is made up of asymmetric and symmetric components, as well as pooling, convolutional, and auxiliary classifiers [ 18 ]. The depthwise separable convolution structure is in the center of the Xception network, mostly made up of entrance flow, middle flow, exit flow, depthwise separable convolution, etc. Convolution and pooling are accomplished by Xception using three flows, where the deep separable convolution that lowers network complexity guarantees the most significant amount of information is sent between layers. At the same time, Xception simultaneously extends the network [ 19 ]. Majority Voting Block In the final block, we introduce a collaborative decision-making approach through a majority voting method. Majority voting is used to enhance prediction performance by utilizing well-trained CNN models. Typically, the majority vote inputs an odd number of models and outputs the results of the voting for those models [ 20 ]. In majority voting, each model provides its estimates individually, and the best-rated class for the final prediction is the class that receives the largest number of votes among all the cast votes [ 21 ]. This decision-making procedure significantly improves our dental image analysis’s overall accuracy while fine-tuning each model’s accuracy. Experimental setup As mentioned earlier, two networks will be trained, the first to locate the teeth and the second to classify them based on their situation. The image dataset used in the training of the YOLOv8 network has been divided into three groups: training, validation, and testing, to achieve the best results. The training set contains 130 images, the verification set includes 50 images, and the test set contains 20 images, with a ratio of 65%, 25%, and 10%, respectively. The YOLOv8n model was used for training; the parameters were determined by the following: image size 640 x 640, the number of training epochs 50, with optimizer SGD, learning rate 0.01, and batch size 16. Careful preparation of the dataset is required before training networks. Due to the lack of available images, we divided the dataset into ten subgroups using the K-fold method. To take advantage of transfer learning, we used Imagenet weights, a set of weights that were pre-trained with 1000 different classes, to save time and effort in the process of training networks from scratch. Each model was trained on the data 10 times, and the network parameters remained constant in each training. Cross-validation is one of the most commonly utilized methods for resampling data to find accurate model prediction errors and fine-tune model parameters. To evaluate a prediction model’s generalizability and prevent overfitting, employ cross-validation [ 22 ]. K-fold cross-validation was used in our research. This method divides the training dataset into subsets of equal size. One subset is chosen as the test dataset for each iteration, while the remaining k-1 subsets are utilized as the training and validation datasets [ 23 ]. We used five different models: VGG19, InceptionResNetV2, InceptionvV3, EfficientNetB0, and Xception. Each model’s output is connected to a flattened layer responsible for representing it to a 1x1 vector. This vector is the summarization of the image features that were extracted by applying the dataset’s images to model layers. To produce an output from this vector, we add three dense layers. The first two layers have 512 and 256 neurons with the ’real’ activation function; each layer is followed by a Dropout function with parameters .4 and .2, respectively. The output layer is the third dense layer that contains one neuron with a ’sigmoid’ activation function. All models had been trained with the same batch size, optimization algorithm, and loss function, which were 5, RMSprop, and ’binary crossentropy’. During model training, we have to control the step size of the model’s weight updating to avoid overshooting the optimal solution. Therefore, we set the learning rate up to 0.00001 for the VGG19 model and 0.0001 for ResNet50V2, Inception-ResNetv2, Xception, and Inception v3 models. Every model had been trained with 90 epochs except the Inception v3 model, which had been trained with 100 epochs. Finally, we defined the metrics parameter in the compile layer as ’accuracy’, which means during model training, the evaluation measures of the model will track and report the accuracy metric. In each iteration, one of the subgroups was considered a test dataset, while the rest were used as training and validation datasets. In the next iteration, another data subgroup was considered a test dataset, while we used the other set as a training and verification dataset. In addition, to ensure that the model starts from the optimal state at each iteration, the model’s hyperparameters were cleared after each training. The Google Colab Pro training environment was used for training CNN models. This environment contains three different GPUs, T4, T100, and A100, for our experimental models. The environment uses an NVIDIA DGX A100 with 40GB of memory and 83.5 GB of system RAM. Performance Measurements A confusion matrix was used to evaluate the models’ performance. In our models, calculating the relevant quantity metrics for a binary classifier could be represented as a 2x2 matrix, as shown in Fig. 6 . This matrix’s elements are True positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN), which denotes the number of positive samples that were correctly classified, the number of negative samples that were correctly classified, the number of positive samples that incorrectly classified and the number of negative samples that incorrectly classified respectively. From the confusion matrix’s elements, we used four methods to check the correctness of the model prediction. First, overall the dataset images, the ratio of correct prediction samples indicates accuracy, the accuracy in Eq. (1). Second, using the ratio of the number of truly predicted samples for one positive class over the summation of truly and falsely the number of predicted samples for the same class. This ratio indicates recall and its Eq. (2). Thirdly, precision, which is determined as the ratio between properly categorized samples and samples allocated to that class, indicates the percentage of the retrieved samples that are relevant as written in Eq. (3). Finally, F1 Score is a harmonic mean of precision and recall as shown in Eq. (4). F1 score is generally used in [ 24 ]. $$\:Accuracy=\frac{TP+TN}{TP+TN+FP+FN}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(1\right)$$ $$\:Recall=\frac{TP}{TP+FN}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(2\right)$$ $$\:Precision=\frac{TP}{TP+FP}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(3\right)\:\:\:$$ $$\:{F}_{1}=2*\frac{Precision*Recall}{Precision+Recall}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(4\right)\:$$ Statistical Analysis To enhance labeling accuracy and minimize bias, two experienced dental radiologists (E.O. and M.T.G.) independently evaluated the images. The images were re-examined using a cross-validation approach. In the Cohen’s Kappa test, this ratio was found to be approxi mately 1.0. This result shows high agreement between the two observers. Ten-fold cross-validation was conducted to evaluate the significance of performance differences between models, calculating mean and standard deviation for critical metrics such as F1-score, sensitivity, specificity, and accuracy. Model performance was primarily measured by accuracy, calculated as a ratio between the summation of TP and TN with the total predicted samples, and additional metrics including precision, sensitivity, and specificity. Due to the non-normal distribution of results, non-parametric tests were employed for statistical analysis. Results This section will cover the networks’ training results as well as our study’s majority voting effectiveness. We used several networks for tooth fracture classification and tooth detection. Regarding the detection performance, the network performed admirably. But the CNN models’ accuracy on their own wasn’t enough. We solved this by enhancing the overall accuracy using a majority vote method. We enhanced the findings by merging the predictions from many models. To highlight our results, let’s title the challenges we will solve: 1) detecting teeth in occlusal teeth images, 2) classifying detected teeth as fractured or non-fractured teeth. For the first challenge, we used the YOLOv8n model for detecting teeth. We evaluated the model by applying 20 images as a test to the model, and it detected the teeth correctly. Figure 7 contains three different images tested by the YOLOv8n model. The model’s performance is measured by the experimentation process’s accuracy and object detection’s robustness. In our model, the mean average precisions (mAPs) are 99.5% and 81.4% for the 50 and (50-95) thresholds, respectively. Figure 8 shows the training results for mAP50 and mAP50-95. Many AI models have been used to train classification networks. Those models were arranged in descending order based on the accuracy of each model, as shown in Table 1. Among all the trained models, VGG19 and InceptionResNetV2 had the highest performance percentages, 87.92% and 87.44%, respectively. The rest received lower performance accuracy, as EfficientNetB0, InceptionV3, VGG16, ResNet50, ResNet50v2, and Xception obtained between 84.65% to 86.93%, which is considered a lower percentage than the previous models. For Recall, Precision, and F1-score evaluation vectors, ResNet50v2, InceptionV3, and VGG19 achieved the highest percentages, which are 92.56%, 92.02%, and 91.01%, respectively. As shown in Table 1, the highest accuracy needed to be higher. Therefore, we used the majority voting method to increase the prediction accuracy. While exploring the best combinations for majority voting, the accuracy of every model was considered. Firstly, we started with the best three models. After that, we moved to a combination of the highest accuracy of four models, until the best combination of seven models was identified. Finally, we used all eight models in one combination. The majority voting method was employed in each combination to determine which arrangement produced the best accuracy. This method of choosing the combination that gave the highest accuracy was applied consistently to all models, providing an analytical evaluation of ensemble performance in various configurations. To enhance the generality of our exploration of model combinations, we decided to use a random selection procedure to ensure the best model combination was picked out of all of our explorations. Table 2 shows the result of the combined model exploration. All eight models are arranged in ascending order at the beginning, shown in the first row. The subsequent rows in the table represent a distinct combination of CNN models, highlighted by light blue blocks indicating the inclusion of a particular model in that combination. Each highlighted block corresponds to the model used in the combination based on the arrangement established in the first table row. The accuracy column on the right displays the majority voting ensemble’s overall accuracy for each model combination. Starting with the three highest model accuracies, which are VGG19, Inception-ResNetV2, and ResNet50v2, we got 90.95%. For the second combination, we added the InceptionV3 model to the previous one, and the accuracy decreased to 89.69%. The accuracy increased to 90.68% after adding the Xception model to the second combination. In the fourth combination, adding VGG16 models decreased the overall accuracy to 90.19%. However, the accuracy is reduced when the ResNet50 model is added. In the last scenario for creating majority voting combination-based models’ accuracy, the majority voting accuracy increased to 91.19% while using all models as one combination. Finally, randomly integrating VGG19, EfficientNetB0, InceptionResNetV2, InceptionV3, and Xception models yielded the highest accuracy of 91.94%. From Table 2, we can conclude that the results demonstrate the effectiveness of majority voting as an ensemble method. The total accuracy is significantly increased using the combined collective decisions of many models. The rising accuracy percentages highlight the benefits of mixing several models, which improves overall performance. Table 3 shows the fusion evaluation vectors. Table 3 shows that the highest accuracy among the five CNN models is VGG19’s accuracy, which is 87.92. In comparison with VGG19 accuracy, the Fusion accuracy is increased to 91.94%. Also, for recall, the fusion recall evaluation is 94.14% compared to the highest five models' recall, which is 92%. Moreover, for precision, the CNN model’s highest precision evaluation is 92%, and the fusion’s precision is 93.41%, which is higher than the CNN model’s precision evaluation. Finally, the fusion f1-score evaluation is 93.77%, which is higher than the highest CNN model, which is 91% for VGG19. We used K-fold cross-validation to evaluate the network models. Datasets are divided into 10 dataset subgroups to ensure the model’s generalization gives us a reliable and robust estimation. Figure 9 is the confusion matrix for the general model; the numbers of fractured (TP) and non-fractured (TN) samples that were predicted correctly are 241 and 15, respectively. On the other hand, the number of fractured samples, but the model predicted as non-fractured (FN), is 15; moreover, 17 samples are predicted as non-fractured and fractured (FP). Although the fusion model demonstrated improved performance across all evaluation metrics, we acknowledge the limitations inherent in training on a relatively small dataset of 398 cropped images. Moreover, given the binary classification task, we prioritized recall and F1-score in our analysis, as these metrics are more clinically meaningful. In particular, maximizing recall reduces the likelihood of false negatives, which is critical in medical diagnostics to avoid missed fracture cases. Discussion Patients who experience dental trauma may face various challenges in obtaining a timely and accurate diagnosis. These challenges can stem from the need for rapid post-trauma assessment, the limited availability of specialized dentists, variations in individual diagnostic experience, and the inherent limitations of radiographic imaging techniques. Although most modern dental clinics are equipped with radiographic imaging systems, the impact of these techniques on diagnostic accuracy should not be overlooked. At this point, AI-based analysis systems have the potential to assist clinicians in dental trauma assessments by standardizing the diagnostic process and enhancing accuracy. Various studies can detect fractures in jaw bones or teeth using AI models. Dong-Min Son et al. [ 25 ], offer automated mandibular fracture detection using DL techniques. This work suggests using DL methods to detect mandibular fractures from a panoramic radiograph without the need for radiologists’ assistance. You Only Look Once (YOLO) is a one-stage detection technique that is the DL system utilized in this work. Panoramic radiographs were pre-processed using gamma modulation, multi-bounding boxes, single-scale luminance adaptation transform (SLAT), and multi-scale luminance adaptation transform (MLAT) to increase the detection accuracy of the system. In the training and test datasets, the fractures’ form and position are not consistent in location and shape. In addition, mandibular fractures in panoramic radiographs show considerable curvature at the background level, and the original panoramic radiographs are dark. Data augmentation and preprocessing approaches mentioned above demonstrated detection accuracy for mandibular fractures with ambiguous forms and regions. Another study conducted by Son et al. [ 26 ] used combined deep-learning techniques to aid in the diagnosis of mandibular fractures. They used combined DL techniques, which are YOLO and U-Net algorithms. Without CBCT, mandibular fractures were detected using the auxiliary diagnostic algorithms YOLO and U-Net based on panoramic images. The purpose of combining these algorithms in this network is that each one has an advantage that the other does not. They used U-net because it helps segment widely dispersed fracture locations. Hence, it is utilized in conjunction with U-net to complement undiagnosed instances of mandibular fracture, while YOLO is weak in this field. At the same time, they used YOLO because of its advantage in detecting fracture locations. Moreover, some studies use AI to detect fractures in the maxillofacial region. Warin et al. [ 27 ] evaluated CT images in 4 classes: frontal fracture, midface fracture, mandibular fracture, and no fracture. Warin et al. [ 28 ] examined midface fractures in CT images, while other studies evaluated fractures in the mandibular bone in panoramic images [ 29 – 31 ]. In Nishiyama et al. [ 32 ], used DL models to detect fractures in the mandibular condyle region in panoramic images. Some studies use DL models to detect tooth root fractures. Some studies can detect vertical root fractures in CBCT images [ 6 , 33 ], or panoramic images [ 34 ] using DL models. Chang et al. [ 35 ] used AI architectures to predict the probability of root fracture after a crown or root canal treatment. Three different CNN architectures were used in a study that detected fractures in dental implants in panoramic and periapical radiographs [ 36 ]. Few studies in the literature have explored the application of AI in the domain of dental traumatology. Notably, Bani-Hani et al. [ 37 ] introduced a CNN-based model using periapical radiographs to classify dental fractures into four subtypes. However, their work lacked automated tooth localization, relied on a relatively small dataset (108 fractured teeth), and reported a moderate overall accuracy of 78.7%. In contrast, the present study, for the first time, leverages OR for fracture detection, employs YOLOv8-based automatic localization of the maxillary central incisors, and applies an ensemble classification approach combining five CNN architectures through majority voting. With a more extensive and balanced dataset (398 teeth), significantly higher classification accuracy (91.94%), and potential for real-time application, the proposed method provides a technically and clinically superior diagnostic tool. These attributes underscore the originality and practical relevance of our approach in comparison to prior literature. In the current study, tooth fractures were detected using OR images. During the training phase of our tooth detection models, we experimented with multiple versions of the YOLO algorithms, including YOLOv8, YOLOv9, and YOLOv10, to assess their effectiveness in accurately detecting tooth localizations. Throughout the process, the training and testing accuracy metrics for these versions were consistently high and showed minimal variance, indicating that models had effectively learned the patterns within the dataset. However, while evaluating these models on test samples, YOLOv8 showed the best robustness and stability, outperforming YOLOv9 and YOLOv10. This robustness behavior in YOLOv8 leads to more generalizability in many different data scenarios, and we selected this model as our primary tooth detection approach. In the subsequent step of the proposed study, the locations of the 11 and 21 teeth, identified using the YOLOv8 detector, were utilized for classification. Contrary to the detection step, no pre-processing was employed during the classification. Various pre-trained classifier networks were employed, and accuracy rates of 84.65% and 87.92% were obtained. To further improve those results, majority voting was also utilized in the study. It allows the ensemble to compensate for the limitations of individual models, capturing a broader range of patterns and achieving better generalization. Through the application of majority voting, the overall accuracy increased by 4.6–91.94%. In the present study, tooth fracture detection was performed on maxillary central incisors using the YOLOv8 algorithm, achieving an accuracy rate of 91.94%. This outcome demonstrates that image-based AI models offer more direct, precise, and clinically applicable solutions for dental trauma diagnosis compared to traditional large language models (LLMs). For instance, Özden et al. [ 38 ] reported that ChatGPT and Google Bard (Gemini), two widely used general-purpose LLMs, provided correct responses to trauma-related dental questions with only 57.5% accuracy, a figure falling below the acceptable threshold for clinical reliability. Similarly, in a study by Tokgöz Kaplan and Cankar [ 39 ] focusing on dental avulsion, the Gemini model outperformed ChatGPT-3.5 in terms of overall accuracy (p < 0.004); however, both models exhibited limited success in responding to multiple-choice technical questions. These findings highlight the inherent limitations of LLMs in clinical decision-making, particularly in domains that require visual interpretation and domain-specific training. In contrast, our use of the YOLOv8 model, trained directly on orthopantomographic (OPG) images, allowed for accurate localization and classification of dental fractures, demonstrating the model’s practical utility as a decision-support tool. Furthermore, the scope of our study extends beyond model architecture to include expert-driven data curation, annotation, and the application of ensemble learning (majority voting) to enhance classification robustness. In this regard, our work distinguishes itself methodologically and technically from prior LLM-based approaches and underscores the potential of image-based deep learning systems to provide more effective and reliable diagnostic support in the field of dental traumatology. Conclusion Accurate detection and classification of dental fractures are crucial for effective diagnosis and treatment planning. In this study, an AI-based automated system was developed for detecting fractures in maxillary central incisors. The proposed method achieved an accuracy rate of 91.94% by utilizing YOLOv8 for tooth detection and CNN-based models for fracture classification. This study is significant as one of the first investigations employing DL techniques for dental fracture detection using OR. The developed AI-based automated approach can assist clinicians in making faster and more consistent diagnoses, reducing interobserver variability. The integration of AI-assisted tools in dentistry holds the potential to enhance diagnostic accuracy and improve patient outcomes. Limitations This study has several limitations. First, only OR was utilized, without validation through other imaging techniques such as periapical radiography or CBCT. As a result, the three-dimensional structure and orientation of fractures could not be fully assessed. Additionally, the dataset was obtained from a single center, limiting the generalizability of the model to different patient populations. Moreover, the study exclusively focused on crown fractures, classifying them as either “fractured” or “non-fractured,” without considering root fractures or other dental anomalies. Finally, the model was trained and tested solely on maxillary central incisors (11 and 21), and its performance on other tooth groups was not evaluated. Considering these limitations, further improvements in detection and classification accuracy are required for the proposed method to be effectively integrated into clinical practice. Future studies should incorporate additional imaging modalities, such as periapical radiography and CBCT, to enhance the assessment of fracture orientation and depth. Furthermore, evaluating the model using larger and more diverse datasets from multiple centers will be crucial for improving its generalizability. Expanding the classification to include both crown and root fractures, along with their subtypes, could provide a more comprehensive diagnostic framework. Additionally, testing the model on different tooth groups would further validate its applicability and diagnostic performance. The originality of this work stems from the problem itself and the comprehensive analysis applied to it, rather than from architectural innovation. Substantial effort was dedicated to the collection, curation, and meticulous labeling of the dataset, which required considerable domain expertise and time. Given the resource-intensive nature of this process, we made a deliberate decision to focus on fine-tuning well-established, high-performing architectures rather than developing new ones, in order to ensure that our results would be robust, reproducible, and directly comparable to the state-of-the-art. Declarations Author Contributions Ahmed N. Nusari: Writing – review & editing, Writing – original draft, Methodology, Investigation, Conceptualization. Esra ÖNCÜ: Writing – review & editing, Supervision, Investigation, Conceptualization. Emin Argun Oral: Writing – original draft, Methodology, Investigation, Conceptualization. İbrahim Yücel Özbek: Writing – original draft, Investigation, Conceptualization. Özkan MİLOĞLU: Writing – original draft, Final approval. Mustafa Taha GÜLLER: Final approval. Funding This research received no external funding. Ethical approval Ethics committee approval was obtained from Atatürk University Faculty of Dentistry Ethics Committee, date: 2023/ number: 61. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request. Informed consent Not applicable Acknowledgments Not applicable Clinical trial number Not applicable References Bourguignon C, Cohenca N, Lauridsen E, Flores MT, O'Connell AC, Day PF, et al. International Association of Dental Traumatology guidelines for the management of traumatic dental injuries: 1. Fractures and luxations. Dent Traumatol. 2020;36(4):314–30. Jogezai U, Kalsi A. Long-term complications and management of dental trauma in the adult patient-Part 1: fractured teeth, pulpal complications and resorption. Br Dent J. 2024;237(2):95–105. Kullman L, Al Sane M. Guidelines for dental radiography immediately after a dento-alveolar trauma, a systematic literature review. Dent Traumatol. 2012;28(3):193–9. Patel S, Puri T, Mannocci F, Navai A. Diagnosis and management of traumatic dental injuries using intraoral radiography and cone-beam computed tomography: an in vivo investigation. J Endod. 2021;47(6):914–23. Huang C, Wang J, Wang S, Zhang Y. A review of deep learning in dentistry. Neurocomputing. 2023;554:126629. Hu Z, Cao D, Hu Y, Wang B, Zhang Y, Tang R, et al. Diagnosis of in vivo vertical root fracture using deep learning on cone-beam CT images. BMC Oral Health. 2022;22(1):382–9. Abdelazim R, Fouad EM. Artificial intelligent-driven decision-making for automating root fracture detection in periapical radiographs. BDJ Open. 2024;10(1):76. Sarkar M, Mandal A. Slaahe: selective apex adaptive histogram equalization. Frankl Open. 2023;3:100023. Terven J, Córdova-Esparza DM, Romero-González JA. A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo-nas. Mach Learn Knowl Extr. 2023;5(4):1680–716. Sary IP, Andromeda S, Armin EU. Performance comparison of YOLOv5 and YOLOv8 architectures in human detection using aerial images. Ultima Computing: J Syst Comput. 2023;15(1):8–13. Paymode AS, Malode VB. Transfer learning for multi-crop leaf disease image classification using convolutional neural network VGG. Artif Intell Agric. 2022;6:23–33. Liu Y, Wang J. Transfer learning based multi-layer extreme learning machine for probabilistic wind power forecasting. Appl Energy. 2022;312:118729. Roy AM. Adaptive transfer learning-based multiscale feature fused deep convolutional neural network for EEG MI multiclassification in brain–computer interface. Eng Appl Artif Intell. 2022;116:105347. Vignesh S, Savithadevi M, Sridevi M, Sridhar R. A novel facial emotion recognition model using segmentation VGG-19 architecture. Int J Inform Technol. 2023;15(4):1777–87. Mohamed ISE, Zarif S, Abdel-Kader H, Elsayed M. Breast cancer detection based on hybrid deep learning models. IJCI Int J Computers Inform. 2023;10(3):119–25. Bhardwaj P, Kumar S, Kumar Y. A comprehensive analysis of deep learning-based approaches for the prediction of gastrointestinal diseases using multi-class endoscopy images. Arch Comput Methods Eng. 2023;30(7):4499–516. Das B, Saha A, Mukhopadhyay S. Rain removal from a single image using refined inception ResNet v2. Circuits Syst Signal Process. 2023;42(6):3485–508. Srinivas K, Gagana Sri R, Pravallika K, Nishitha K, Polamuri SR. COVID-19 prediction based on hybrid Inception V3 with VGG16 using chest X-ray images. Multimedia Tools Appl. 2024;83(12):36665–82. Liu Y, Zhang L, Hao Z, Yang Z, Wang S, Zhou X, et al. An xception model based on residual attention mechanism for the classification of benign and malignant gastric ulcers. Sci Rep. 2022;12(1):15365. Jhang K. Voting and ensemble schemes based on CNN models for photo-based gender prediction. J Inform Process Syst. 2020;16(4):809–19. Jain A, Kumar A, Susan S. Evaluating deep neural network ensembles by majority voting cum meta-learning scheme. Springer. 2022;29–37. Berrar D, Cross-Validation. Encyclopedia Bioinf Comput Biology. 2019;1:542–5. Alaoui SS, Labsiv Y, Aksasse B. Classification algorithms in data mining. Int J Tomography Simul. 2018;31(4):34–44. Hicks SA, Strümke I, Thambawita V, Hammou M, Riegler MA, Halvorsen P, et al. On evaluation metrics for medical applications of artificial intelligence. Sci Rep. 2022;12(1):5979. Son DM, Yoon YA, Kwon HJ, An CH, Lee SH. Automatic detection of mandibular fractures in panoramic radiographs using deep learning. Diagnostics. 2021;11(6):933. Son DM, Yoon YA, Kwon HJ, Lee SH. Combined deep learning techniques for mandibular fracture diagnosis assistance. Life. 2022;12(11):1711. Warin K, Limprasert W, Suebnukarn S, Paipongna T, Jantana P, Vicharueang S. Maxillofacial fracture detection and classification in computed tomography images using convolutional neural network-based models. Sci Rep. 2023;13(1):3434. Warin K, Vicharueang S, Jantana P, Limprasert W, Thanathornwong B, Suebnukarn S. Deep learning for midfacial fracture detection in CT images. MEDINFO 2023-The Future Is Accessible. IOS; 2024. pp. 1497–8. Vinayahalingam S, van Nistelrooij N, van Ginneken B, Bressem K, Tröltzsch D, Heiland M, et al. Detection of mandibular fractures on panoramic radiographs using deep learning. Sci Rep. 2022;12(1):19596. Yari A, Fasih P, Hosseini Hooshiar M, Goodarzi A, Fattahi SF. Detection and classification of mandibular fractures in panoramic radiography using artificial intelligence. Dentomaxillofacial Radiol. 2024;53(6):363–71. Warin K, Limprasert W, Suebnukarn S, Inglam S, Jantana P, Vicharueang S. Assessment of deep convolutional neural network models for mandibular fracture detection in panoramic radiographs. Int J Oral Maxillofac Surg. 2022;51(11):1488–94. Nishiyama M, Ishibashi K, Ariji Y, Fukuda M, Nishiyama W, Umemura M, et al. Performance of deep learning models constructed using panoramic radiographs from two hospitals to diagnose fractures of the mandibular condyle. Dentomaxillofacial Radiol. 2021;50(7):20200611. Yang P, Guo X, Mu C, Qi S, Li G. Detection of vertical root fractures by cone-beam computed tomography based on deep learning. Dentomaxillofacial Radiol. 2023;52(3):20220345. Fukuda M, Inamoto K, Shibata N, Ariji Y, Yanashita Y, Kutsuna S, et al. Evaluation of an artificial intelligence system for detecting vertical root fracture on panoramic radiography. Oral Radiol. 2020;36:337–43. Chang WT, Huang HY, Lee TM, Sung TY, Yang CH, Kuo YM. Predicting root fracture after root canal treatment and crown installation using deep learning. J Dent Sci. 2024;19(1):587–93. Lee DW, Kim SY, Jeong SN, Lee JH. Artificial intelligence in fractured dental implant detection and classification: evaluation using dataset from two dental hospitals. Diagnostics. 2021;11(2):233. Bani-Hani T, Wedyan M, Al-Fodeh R, Shuqeir R, Al Jundi S, Tewari N. Artificial intelligence model for application in dental traumatology. Eur Arch Paediatr Dent. 2025 May;31. 10.1007/s40368-025-01063-0 . Ozden I, Gokyar M, Ozden ME, Sazak Ovecoglu H. Assessment of artificial intelligence applications in responding to dental trauma. Dent Traumatol. 2024;40(6):722–9. Tokgöz Kaplan T, Cankar M. Evidence-based potential of generative artificial intelligence large language models on dental avulsion: ChatGPT versus Gemini. Dent Traumatol. 2025;41(2):178–86. Tables Tables 1 to 3 are available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files Tables.docx Cite Share Download PDF Status: Published Journal Publication published 18 Dec, 2025 Read the published version in BMC Oral Health → Version 1 posted Editorial decision: Revision requested 10 Oct, 2025 Reviews received at journal 07 Oct, 2025 Reviewers agreed at journal 06 Oct, 2025 Reviews received at journal 06 Oct, 2025 Reviewers agreed at journal 06 Oct, 2025 Reviewers agreed at journal 05 Oct, 2025 Reviews received at journal 30 Sep, 2025 Reviewers agreed at journal 26 Sep, 2025 Reviewers agreed at journal 22 Sep, 2025 Reviewers invited by journal 21 Sep, 2025 Editor invited by journal 16 Sep, 2025 Editor assigned by journal 03 Aug, 2025 Submission checks completed at journal 03 Aug, 2025 First submitted to journal 24 Jul, 2025 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-7207126","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":522656986,"identity":"fe96a1c3-8b8d-42d5-93f1-d1829f80bce4","order_by":0,"name":"Ahmed NUSARI","email":"","orcid":"","institution":"Ataturk University","correspondingAuthor":false,"prefix":"","firstName":"Ahmed","middleName":"","lastName":"NUSARI","suffix":""},{"id":522656988,"identity":"bfdcb4b6-7031-4f26-87b0-927a053cd31b","order_by":1,"name":"Esra ONCU","email":"","orcid":"","institution":"Ataturk University","correspondingAuthor":false,"prefix":"","firstName":"Esra","middleName":"","lastName":"ONCU","suffix":""},{"id":522656990,"identity":"51857ecf-77a9-417e-8167-5d656c6607fc","order_by":2,"name":"Emin Argun ORAL","email":"","orcid":"","institution":"Ataturk University","correspondingAuthor":false,"prefix":"","firstName":"Emin","middleName":"Argun","lastName":"ORAL","suffix":""},{"id":522656992,"identity":"a8ce5876-0960-4c07-95a5-3094067cef19","order_by":3,"name":"Ibrahim Yucel OZBEK","email":"","orcid":"","institution":"Ataturk University","correspondingAuthor":false,"prefix":"","firstName":"Ibrahim","middleName":"Yucel","lastName":"OZBEK","suffix":""},{"id":522656993,"identity":"ee6f9333-f6eb-4950-8abe-7cd2625f14e9","order_by":4,"name":"Ozkan MILOGLU","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA10lEQVRIiWNgGAWjYFACxsYDDAwWDPwSYJ6EDDFaGoBaJBgkZwBZQAYPUfaAtRjcAGthIKyFX+xww4EPvyQSN99uPv7oRo0FDwP74aMb8GmRnJ3YcHBmn0TitjvHEptzjgEdxpOWdgOfFoPbiQ2HeXuAWm7kGDbnsAG1SPCYEadl8wyQln/EauH5IZG4QQKoJbeNCC0QvzRIGM+4kZY4O7dPgoeNkF/4pdMfPvjwx0a2f0bygc853+rk+NkPH8OrBQwY25A4bASVg8Ef4pSNglEwCkbBCAUAl3RN+eBQab4AAAAASUVORK5CYII=","orcid":"","institution":"Ataturk University","correspondingAuthor":true,"prefix":"","firstName":"Ozkan","middleName":"","lastName":"MILOGLU","suffix":""},{"id":522656995,"identity":"672355e8-0cc2-457e-b81d-a4dc82bf6e89","order_by":5,"name":"Mustafa Taha GULLER","email":"","orcid":"","institution":"Giresun University","correspondingAuthor":false,"prefix":"","firstName":"Mustafa","middleName":"Taha","lastName":"GULLER","suffix":""}],"badges":[],"createdAt":"2025-07-24 15:38:16","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7207126/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7207126/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s12903-025-07445-6","type":"published","date":"2025-12-18T15:57:12+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":92682615,"identity":"64041420-ac9e-4fd7-a2f9-9f32be0e8fd3","added_by":"auto","created_at":"2025-10-03 01:19:14","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":654526,"visible":true,"origin":"","legend":"","description":"","filename":"revisedmanuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/d90690372993b90a91cc6a35.docx"},{"id":92682613,"identity":"2980b815-ddaa-4cdb-ae57-508016bb8fc7","added_by":"auto","created_at":"2025-10-03 01:19:14","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7701,"visible":true,"origin":"","legend":"","description":"","filename":"9c86e3739836456295b30e7cd18fa0c5.json","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/798a9506922462e70c3b199f.json"},{"id":92682095,"identity":"103b1853-b863-415f-90ef-f7d475ec6c39","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":122699,"visible":true,"origin":"","legend":"","description":"","filename":"9c86e3739836456295b30e7cd18fa0c51enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/292d0491e1271103025a8302.xml"},{"id":92682115,"identity":"e33de09c-03e5-40a8-96f2-a1641d6a64af","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"jpeg","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":711016,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/b1c09b5b6c6c1084d3a8d84d.jpeg"},{"id":92682113,"identity":"584fc9b9-0057-43ed-b0ab-c8056dee1330","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"jpeg","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1074,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/06e49c270c8a93e95982fafd.jpeg"},{"id":92682620,"identity":"ca54cb02-4a90-4cad-9e3b-8073f342f95a","added_by":"auto","created_at":"2025-10-03 01:19:14","extension":"jpeg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":210217,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage11.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/0c52b219d52f361bab61823b.jpeg"},{"id":92682616,"identity":"88e9024c-33c0-419c-8001-d52162a8fc8a","added_by":"auto","created_at":"2025-10-03 01:19:14","extension":"jpeg","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1074,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/88a29b2d89f2af302eb90012.jpeg"},{"id":92682104,"identity":"ad65988d-b2cd-42e5-a148-a5900698f46a","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"jpeg","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":847306,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage13.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/94ad2f4ed0873e2a1fe9a96f.jpeg"},{"id":92682129,"identity":"6f46598b-a2a0-4baa-bde8-f2777a618a7e","added_by":"auto","created_at":"2025-10-03 01:11:15","extension":"jpeg","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1074,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/c2fbb5465d33932c36757454.jpeg"},{"id":92683109,"identity":"f65f6505-7cc6-412f-96a5-6c8d9daecd60","added_by":"auto","created_at":"2025-10-03 01:35:15","extension":"jpeg","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":308848,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage15.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/0a7aeae5748666af3565903c.jpeg"},{"id":92682112,"identity":"20ef9c44-87f3-4a47-85c0-977fe5e85667","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"jpeg","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1074,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/a8eb221ba0a6da6492af769b.jpeg"},{"id":92682618,"identity":"191129e4-3ee2-4653-92e6-3fc503492f59","added_by":"auto","created_at":"2025-10-03 01:19:14","extension":"jpeg","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":256207,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage17.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/4b7f1af92623395c49fd27cd.jpeg"},{"id":92682617,"identity":"fdb190d2-9ba0-408c-9895-41bf892c7518","added_by":"auto","created_at":"2025-10-03 01:19:14","extension":"jpeg","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1074,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/ba3e16bb4b3f5ff4c0a59803.jpeg"},{"id":92682128,"identity":"552909d5-029b-4454-b4c7-5dad582e7c3e","added_by":"auto","created_at":"2025-10-03 01:11:15","extension":"jpeg","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":462301,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/c5e00fd6404214f3f543879e.jpeg"},{"id":92682102,"identity":"3e136ad6-7903-49ce-8e07-e08e52ab866d","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"jpeg","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1074,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/a1d5bc859b0be3dcba0ae0f5.jpeg"},{"id":92682140,"identity":"50597110-5d9c-497f-9fc6-cf728dd7011b","added_by":"auto","created_at":"2025-10-03 01:11:15","extension":"jpeg","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1115722,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/4e3e2a294eb0c017fc801fef.jpeg"},{"id":92682114,"identity":"c0418b08-a663-4b76-a055-3bd7162334ef","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"jpeg","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1074,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/3f03822f721a49c704ce6c91.jpeg"},{"id":92682624,"identity":"d0e0717a-37ea-41a3-8865-2f1732cc07c8","added_by":"auto","created_at":"2025-10-03 01:19:15","extension":"jpeg","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":603379,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/061c79e545ba3a9d2d7a9421.jpeg"},{"id":92682628,"identity":"654a057f-d594-4215-a688-8ed66d38c22d","added_by":"auto","created_at":"2025-10-03 01:19:15","extension":"jpeg","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1074,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/ff85e70d13b6946d90706f8c.jpeg"},{"id":92682999,"identity":"22962e6b-3a29-4fe7-ac68-123d207835b2","added_by":"auto","created_at":"2025-10-03 01:27:14","extension":"jpeg","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":184629,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage9.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/5fb78868d412217d1925ee4e.jpeg"},{"id":92682623,"identity":"c09ce2d4-43ea-4270-a168-f50ef8fe48d5","added_by":"auto","created_at":"2025-10-03 01:19:14","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":137133,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/445826ba42aeb8e11f62df48.png"},{"id":92682123,"identity":"6d2410cd-d5ed-40d2-8400-a366b9f12b18","added_by":"auto","created_at":"2025-10-03 01:11:15","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":935,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/45fbfb54cf5921baabd8437a.png"},{"id":92682106,"identity":"002db956-29eb-4421-ad81-40e7bfcc0a08","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":31991,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/c608e57efc5ccea603042111.png"},{"id":92682108,"identity":"7fa69d0a-8644-4f05-a09b-ae4f52856f4a","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":935,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/5791cb8d76d52553ee3537b7.png"},{"id":92682124,"identity":"f1e6b9de-22ae-41e3-874b-84a8b565553a","added_by":"auto","created_at":"2025-10-03 01:11:15","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":218847,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/c6e92d292b15357fd135e5cc.png"},{"id":92682626,"identity":"24a373c1-069b-41a3-a211-7e96540a19f3","added_by":"auto","created_at":"2025-10-03 01:19:15","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":935,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/dc447204dcbfa67275c1e616.png"},{"id":92682622,"identity":"2e651b3e-49eb-450e-8b09-0a866000e1b5","added_by":"auto","created_at":"2025-10-03 01:19:14","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":93189,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage15.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/a3d7c7d9f1c22f05ad3d563b.png"},{"id":92682110,"identity":"c471ab40-8818-4009-b076-17280d36eb59","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":935,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/9c7f7f98680b52864067bc9c.png"},{"id":92682122,"identity":"8e473639-0ff4-4261-a6dd-485103fc5829","added_by":"auto","created_at":"2025-10-03 01:11:15","extension":"png","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":45678,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage17.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/8ebe9a5d6cc17b39373693b3.png"},{"id":92682120,"identity":"b98e8608-a3cd-4c15-b406-9b371e468872","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"png","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":935,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/955c0a7f1ceb8e721cfe1228.png"},{"id":92682632,"identity":"9817b7e9-494a-46d1-be7e-6e291bfe3c94","added_by":"auto","created_at":"2025-10-03 01:19:15","extension":"png","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":79368,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/d63297a45b20b7f2326b9775.png"},{"id":92682135,"identity":"58a52808-a8a4-4799-adf7-2f3c9a169cbf","added_by":"auto","created_at":"2025-10-03 01:11:15","extension":"png","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":935,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/a0781f7789e3014d0a902c63.png"},{"id":92682131,"identity":"dad679b4-6623-440f-b6b6-c717da14cc50","added_by":"auto","created_at":"2025-10-03 01:11:15","extension":"png","order_by":32,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":268496,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/cdaa4027c81d2a26a846c800.png"},{"id":92683001,"identity":"fc2578dd-f692-468a-83c5-475f52107a85","added_by":"auto","created_at":"2025-10-03 01:27:15","extension":"png","order_by":33,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":935,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/97e7101711964f100a1bd204.png"},{"id":92682116,"identity":"ae8f4b75-ac58-460e-830a-15670193667b","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"png","order_by":34,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":202501,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/77a2b239e30387dcbfd5c648.png"},{"id":92682137,"identity":"7e040088-cbfc-41a6-a8e0-56e63e4c88a4","added_by":"auto","created_at":"2025-10-03 01:11:15","extension":"png","order_by":35,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":935,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/c66cb46e36fd29a40a01a33b.png"},{"id":92682133,"identity":"42eeadc1-583c-433f-8b43-7fbd935c394c","added_by":"auto","created_at":"2025-10-03 01:11:15","extension":"png","order_by":36,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":24189,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/5c2efbc9ae81e0bb632541ec.png"},{"id":92682127,"identity":"9d557c04-c3ad-4144-aea4-2cd6ca74adcc","added_by":"auto","created_at":"2025-10-03 01:11:15","extension":"xml","order_by":37,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":123146,"visible":true,"origin":"","legend":"","description":"","filename":"9c86e3739836456295b30e7cd18fa0c51structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/3193d47f0fa5534a89f8c587.xml"},{"id":92682627,"identity":"1167c34f-fb6a-4fa9-ac0e-b60f4f073d31","added_by":"auto","created_at":"2025-10-03 01:19:15","extension":"html","order_by":38,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":133129,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/b51692ed31e1845d56473b18.html"},{"id":92682088,"identity":"f1f53e49-d0e3-4055-8300-8055dd0adefa","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":430560,"visible":true,"origin":"","legend":"\u003cp\u003eDataset Images: a) Non-Fractured Teeth, b-f) Fracture Teeth\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/83205a102674c9111c33073f.png"},{"id":92682091,"identity":"1d8695b7-e20e-4eab-bade-503768389fa2","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":232963,"visible":true,"origin":"","legend":"\u003cp\u003ea) Original OR Image b) Labeled OR Image.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/944492b81c338b206f19373e.png"},{"id":92682093,"identity":"5712c31a-9584-40f4-96bd-d61ef4aabe09","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":428178,"visible":true,"origin":"","legend":"\u003cp\u003eGeneral Structure\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/2bc325ebfe02436484de2cab.png"},{"id":92682101,"identity":"e9554e1d-4f77-477a-8fcf-94cee5dc9178","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":219616,"visible":true,"origin":"","legend":"\u003cp\u003ea) Original Occlusal Radiography Image b) Processed Occlusal Radiography Image.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/71abd4bb07b8014f594f05f8.png"},{"id":92682621,"identity":"603be32f-ff6f-4bab-915a-e6bd2b3b36e8","added_by":"auto","created_at":"2025-10-03 01:19:14","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":26900,"visible":true,"origin":"","legend":"\u003cp\u003ea) 11 Teeth b) 21 Teeth.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/3df233d47ac36ed3abfd9bf2.png"},{"id":92682121,"identity":"dee174d2-ae7c-45b5-bc4d-aed5550fc690","added_by":"auto","created_at":"2025-10-03 01:11:15","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":68051,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion Matrix.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/e54bed0b8c9ab08aea66a564.png"},{"id":92683000,"identity":"5a979e01-f673-4978-8e93-a80267d02fd7","added_by":"auto","created_at":"2025-10-03 01:27:15","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":621746,"visible":true,"origin":"","legend":"\u003cp\u003eYolov8 tested images.\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/cf8231b6445a0d93028d9d3d.png"},{"id":92682097,"identity":"7a60788f-1b57-47bd-99ac-058d34b8c39d","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":146800,"visible":true,"origin":"","legend":"\u003cp\u003emAP50 and mAP50-95 Training Result.\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/c2c66d0f2de4064f2510c989.png"},{"id":92682099,"identity":"f05eba48-fbd1-4a18-b5df-c45ff900a8d8","added_by":"auto","created_at":"2025-10-03 01:11:14","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":102570,"visible":true,"origin":"","legend":"\u003cp\u003eMajority Voiting’s Confusion Matrix.\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/12f63672f3a62e092e5d4438.png"},{"id":98813851,"identity":"90f86b0d-1700-4b3a-8582-4c972ca75816","added_by":"auto","created_at":"2025-12-22 16:05:23","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3084993,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/b2b1ac96-8767-4590-b9cc-b493c3af34a9.pdf"},{"id":92682998,"identity":"2ee8a4a7-eaf5-4ef6-835e-7af07c231692","added_by":"auto","created_at":"2025-10-03 01:27:14","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":17906,"visible":true,"origin":"","legend":"","description":"","filename":"Tables.docx","url":"https://assets-eu.researchsquare.com/files/rs-7207126/v1/ff14fc8f40ef2db61d8e0d3f.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Deep Learning-Based Fractured Tooth Detection in Occlusal Radiographs","fulltext":[{"header":"Introduction","content":"\u003cp\u003eTraumatic dental injuries constitute a significant concern in dentistry, as they can lead to functional, aesthetic, and psychological complications. These injuries are particularly prevalent in maxillary central incisors, which hold critical importance in both functional and aesthetic aspects [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Dental trauma may result in crown or root fractures independently, or in some cases, both types of fractures may occur simultaneously. Early diagnosis and appropriate management through suitable treatment modalities are crucial factors in determining the prognosis of the affected tooth [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eRadiographic evaluation should be conducted alongside clinical examination in the diagnosis of dental trauma, with careful assessment of the fracture's size, location, and the extent of damage to the supporting tissues. Periapical, panoramic, and occlusal radiographs (OR) are commonly used imaging modalities in clinical practice for diagnosing dental fractures [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. In addition to these, cone-beam computed tomography (CBCT) is an advanced imaging technique that provides three-dimensional evaluation, particularly in complex fracture cases. OR serves as a crucial diagnostic tool in the assessment of dental trauma and is particularly preferred in cases where periapical radiographs are insufficient or when imaging from additional angles is required [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. The International Association of Dental Traumatology recommends obtaining at least two periapical radiographs and one OR for the evaluation of dental trauma cases, along with the examination of radiographic images taken at different vertical and horizontal angulations [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eIn recent years, artificial intelligence (AI)-powered technologies have made significant advancements in the field of medical image analysis and have increasingly been integrated into dental practice [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Deep learning (DL), a subcategory of AI, refers to sophisticated neural networks composed of multiple layers. Convolutional neural networks (CNNs) represent a specific type of neural network. Due to the capability of CNNs to diagnose and analyze medical images, researchers widely employ them in contemporary studies. This capability stems from the convolutional process, which involves extracting a substantial amount of pixel-level information from images, enabling the model to learn and recognize image patterns effectively [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eTwo fundamental factors play a crucial role in the interpretation of radiological images: the inclusion of diagnostic information within the image and the ability of the evaluator to accurately perceive this information. However, limitations in clinical experience among dentists or a lack of familiarity with previously unencountered cases may lead to incorrect or incomplete diagnoses [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. At this point, the integration of AI-based systems holds the potential to support the diagnostic process by reducing clinician error rates and shortening the time required for diagnosis.\u003c/p\u003e\u003cp\u003eThis study aims to detect crown fractures in maxillary central incisors using an AI-assisted system based on OR. The proposed method consists of two stages: In the first stage, teeth are detected using the YOLOv8 model. In the second stage, the presence of crown fractures is classified using five different CNN-based transfer learning models. Following model evaluations, the majority voting method is applied to enhance diagnostic accuracy. This technique leverages a collective decision-making mechanism among multiple models, minimizing errors while improving overall accuracy.\u003c/p\u003e\u003cp\u003eAccording to the literature review, this study is one of the first investigations on AI-assisted crown fracture detection using OR. Therefore, the findings have the potential to serve as a significant decision-support mechanism for clinicians in both emergency and routine dental practice.\u003c/p\u003e"},{"header":"Materials and methods","content":"\u003cp\u003e\u003cb\u003eData\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThis study protocol was approved by the Ethics Committee of Atat\u0026uuml;rk University Faculty of Dentistry, Department of Oral and Maxillofacial Radiology, Erzurum (Approval No: 2023:61). OR images from 200 patients who visited between 2020 and 2022, routine clinical examination, or follow-up between 2020 and 2022 were analyzed and recorded in PNG format by a clinician.\u003c/p\u003e\u003cp\u003eAll OR images were obtained using the bisecting angle technique with a Belmont PHOT-xIIs Intraoral X-Ray unit (Takara Belmont, Somerset, MA, USA) and recorded on Size 4 (57\u0026times;76 mm) phosphor plates (D\u0026uuml;rr Dental, Bietigheim-Bissingen, Germany). The acquired images were scanned and digitized using the Vistascan Mini Easy System (D\u0026uuml;rr Dental, Bietigheim-Bissingen, Germany). The study included cases with crown fractures in maxillary central incisors (11 and 21 according to the FDI numbering system). However, images that hindered diagnosis due to superimposition or artifacts, teeth with caries, microcrack or dental anomalies, patients with orthodontic appliances were excluded from the study. The teeth were classified as either \u0026ldquo;fractured\u0026rdquo; or \u0026ldquo;non-fractured.\u0026rdquo; Due to the avulsion of two teeth, a total of 398 teeth were evaluated: 258 were classified as fractured and 140 as non-fractured (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows several images from our dataset that contain various fractured teeth. Fractured areas are denoted by orange arrows.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eData annotation\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe labeling process was conducted using LabelImg, a software commonly utilized in DL applications. Each tooth was manually annotated on OR images and classified as either \"fractured\" or \"non-fractured.\"\u003c/p\u003e\u003cp\u003eThe primary objective of this labeling process was to create a high-quality dataset for training the YOLOv8 network. The annotated images were used to enable the model to automatically detect maxillary central incisors (11 and 21). Additionally, the labeled data facilitated the automatic detection of teeth from the original radiographic images and optimized their presentation for the classification phase. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e illustrates the labeling process and how the annotated images were integrated into the AI training workflow.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eThe proposed method\u003c/b\u003e\u003c/p\u003e\u003cp\u003eFor providing optimal patient care, the accurate identification of fractured teeth is essential. A more efficient and objective technique is required due to the growing need for effective dental diagnostics, as manual categorization frequently results in variances in diagnosis. In this proposed method, our primary objective could be obtaining high accuracy in the automated detection and classification of different types of tooth fractures, reducing reliance on humans in diagnosing and classifying teeth, and providing a highly accurate, repeatable, and rapid system for consistent fracture assessments. In the following sections, we will introduce the details of our method by providing its general architecture.\u003c/p\u003e\u003cp\u003eTwo networks are used, and every network has a different purpose. The first network is for detecting teeth, and the other one is for classifying whether they are fractured or not. Different image processing techniques will be applied for each network. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e illustrates the general structure of the steps for detecting and classifying teeth. A tooth detection and classification networks are trained separately. Both networks are combined and used together later. In our approach to dental image analysis, our proposed method unfolds across four integral blocks: Contrast Limited Adaptive Histogram Equalization (CLAHE) / Cropping, detection, classification, and majority voting blocks. Each contributes to a refined and accurate identification of teeth. This contribution gives us reliable results using several intelligence models. By combining these two networks, we can accurately detect and classify teeth, allowing practical analysis and diagnosis of dental conditions.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eCLAHE / Cropping block\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe initial block encompasses two image processing techniques. The first technique, CLAHE, prepares the input image for YOLOv8 training. The CLAHE method improves image quality and enhances color contrast in images by adjusting the relative resolution of lighting in image areas. Thus, the quality of the images can be improved, and the contours of the teeth can be better highlighted, which facilitates the process of accurately and effectively detecting the location of the teeth [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eTo improve the images and highlight the teeth\u0026rsquo;s contour, we applied the CLAHE method twice with varying parameters. The same CLAHE parameters are used for those two times, which are three contrast limits (clipLimit) and (45, 45) of tiles (tileGridSize) as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eCropping the tooth from the original image is the second method. Following YOLOv8\u0026rsquo;s tooth detection, the detected teeth\u0026rsquo;s coordinates are obtained, and a bounding box is added to the original image to point to the detected tooth. The tooth is then cropped according to the box that is painted. As a result of the cropping process, 258 fractured teeth and 140 non-fractured teeth were obtained. Due to size differences, cropped teeth were adjusted to size (448, 224, 3) to guarantee that they could be trained using CNN models as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. This resizing step is essential to maximize the dental classification network training.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eDetection Block\u003c/b\u003e\u003c/p\u003e\u003cp\u003eIn the second block, we used YOLOv8 to provide a foundational understanding of the tooth landscape within the images and discern the distinct features of the 11 and 21 teeth. Based on this detection process, a methodical process was designed to organize the acquisition of precise tooth coordinates precisely determined by the results of YOLOv8. With the help of these obtained coordinates, individual teeth are smoothly repositioned on the original images, and cropping is performed.\u003c/p\u003e\u003cp\u003eUltralytics released YOLOv8 in January 2023. YOLOv8 processes regression, objectness, and classification tasks independently using an anchor-free model with a decoupled head [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. YOLOv8 includes enhancements in the form of a new neural network architecture. This architecture contains two neural networks, the Feature Pyramid Network (FPN) and the Path Aggregation Network (PAN), built on a novel labeling tool that streamlines the annotation process. Several useful features of this labeling tool are adjustable hotkeys, automated labeling, and shortcut labeling. These characteristics together make it simple to annotate images for model training [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThe following image processing was applied to the YOLOv8 network. To detect and locate teeth, the images used in the dataset require resizing because they have different dimensions. To use these images, they have been resized to 640 x 640 so that they are suitable for this process.\u003c/p\u003e\u003cp\u003e\u003cb\u003eClassification Block\u003c/b\u003e\u003c/p\u003e\u003cp\u003eAs we progress to the third block, our focus turns to harnessing CNN networks\u0026rsquo; capabilities. These specialized models extract intricate features and classify teeth based on learned patterns. In this process, different transfer learning is performed using pre-trained CNN models.\u003c/p\u003e\u003cp\u003eThe transfer learning pre-trained CNN was tuned for one task, and its knowledge was transferred to multiple models [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Transfer learning is a successful approach for increasing neural network models\u0026rsquo; capacity to detect and apply previously learned information and abilities to new models [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. It is generally understood that optimizing and thoroughly training a model is a tough and time-consuming process that necessitates using a powerful GPU and millions of training instances. Learning to move to utilize pre-trained weights from various objective circumstances can be a useful technique for overcoming these issues [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eVGG-19 has 19 layers, 16 convolutional layers, 3 Fully Connected layers, and five pooling layers. Batch normalization is utilized after each convolutional layer to alleviate the internal covariate shift problem and make the network quicker and more stable. Except for the last Fully Connected (FC) layer, all layers depend on the Rectified Linear Unit (ReLU) activation function. The filters are always 3x3 in size with a stride of 1. The image\u0026rsquo;s spatial resolution is preserved by using padding of size 1. Max pooling is done throughout a 2x2 window with a stride size of 2 [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. There are 28 layers in the EfficientNetB0: 10 pooling layers, five fully connected layers, and 13 convolutional layers. The convolutional layers use A depthwise separable convolution, which lowers the amount of parameters and processing needed. Max pooling is a straightforward and effective procedure used by the pooling layers. The output from each branch is concatenated, routed through two dense layers, and then passed via a SoftMax activation function for classification. The model is trained using the Adam optimizer and the binary cross-entropy loss function [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eInceptionResNetV2 includes 164 layers and an architecture containing global average pooling 2d, a dense layer, batch normalization, a dropout layer, an activation function, and a second dense layer to categorize the image. InceptionResNetV2 combines feature extraction from Inception with residual connections from ResNet to improve network convergence and gradient flow [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Convolution filters with various kernel sizes (1x1, 3x3, 5x5, and 7x7) and max pooling are used in the module. Concatenating the data enables multi-level feature extraction, which improves network performance. Later, the Inception module was merged with residual connections to generate IR v2, resulting in a more efficient and accurate model [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eInception V3 is a CNN-based classification network. It makes use of inception modules, which have 42 layers deep and are made up of a stacked layer with 1x1, 3 x 3, and 5x5 convolutions. Reduce the number of parameters while increasing the rate of training. This model is made up of asymmetric and symmetric components, as well as pooling, convolutional, and auxiliary classifiers [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThe depthwise separable convolution structure is in the center of the Xception network, mostly made up of entrance flow, middle flow, exit flow, depthwise separable convolution, etc. Convolution and pooling are accomplished by Xception using three flows, where the deep separable convolution that lowers network complexity guarantees the most significant amount of information is sent between layers. At the same time, Xception simultaneously extends the network [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e\u003cb\u003eMajority Voting Block\u003c/b\u003e\u003c/p\u003e\u003cp\u003eIn the final block, we introduce a collaborative decision-making approach through a majority voting method. Majority voting is used to enhance prediction performance by utilizing well-trained CNN models. Typically, the majority vote inputs an odd number of models and outputs the results of the voting for those models [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. In majority voting, each model provides its estimates individually, and the best-rated class for the final prediction is the class that receives the largest number of votes among all the cast votes [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. This decision-making procedure significantly improves our dental image analysis\u0026rsquo;s overall accuracy while fine-tuning each model\u0026rsquo;s accuracy.\u003c/p\u003e\u003cp\u003e\u003cb\u003eExperimental setup\u003c/b\u003e\u003c/p\u003e\u003cp\u003eAs mentioned earlier, two networks will be trained, the first to locate the teeth and the second to classify them based on their situation. The image dataset used in the training of the YOLOv8 network has been divided into three groups: training, validation, and testing, to achieve the best results. The training set contains 130 images, the verification set includes 50 images, and the test set contains 20 images, with a ratio of 65%, 25%, and 10%, respectively. The YOLOv8n model was used for training; the parameters were determined by the following: image size 640 x 640, the number of training epochs 50, with optimizer SGD, learning rate 0.01, and batch size 16.\u003c/p\u003e\u003cp\u003eCareful preparation of the dataset is required before training networks. Due to the lack of available images, we divided the dataset into ten subgroups using the K-fold method. To take advantage of transfer learning, we used Imagenet weights, a set of weights that were pre-trained with 1000 different classes, to save time and effort in the process of training networks from scratch. Each model was trained on the data 10 times, and the network parameters remained constant in each training. Cross-validation is one of the most commonly utilized methods for resampling data to find accurate model prediction errors and fine-tune model parameters. To evaluate a prediction model\u0026rsquo;s generalizability and prevent overfitting, employ cross-validation [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. K-fold cross-validation was used in our research. This method divides the training dataset into subsets of equal size. One subset is chosen as the test dataset for each iteration, while the remaining k-1 subsets are utilized as the training and validation datasets [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. We used five different models: VGG19, InceptionResNetV2, InceptionvV3, EfficientNetB0, and Xception. Each model\u0026rsquo;s output is connected to a flattened layer responsible for representing it to a 1x1 vector. This vector is the summarization of the image features that were extracted by applying the dataset\u0026rsquo;s images to model layers. To produce an output from this vector, we add three dense layers. The first two layers have 512 and 256 neurons with the \u0026rsquo;real\u0026rsquo; activation function; each layer is followed by a Dropout function with parameters .4 and .2, respectively. The output layer is the third dense layer that contains one neuron with a \u0026rsquo;sigmoid\u0026rsquo; activation function.\u003c/p\u003e\u003cp\u003eAll models had been trained with the same batch size, optimization algorithm, and loss function, which were 5, RMSprop, and \u0026rsquo;binary crossentropy\u0026rsquo;. During model training, we have to control the step size of the model\u0026rsquo;s weight updating to avoid overshooting the optimal solution. Therefore, we set the learning rate up to 0.00001 for the VGG19 model and 0.0001 for ResNet50V2, Inception-ResNetv2, Xception, and Inception v3 models. Every model had been trained with 90 epochs except the Inception v3 model, which had been trained with 100 epochs. Finally, we defined the metrics parameter in the compile layer as \u0026rsquo;accuracy\u0026rsquo;, which means during model training, the evaluation measures of the model will track and report the accuracy metric.\u003c/p\u003e\u003cp\u003eIn each iteration, one of the subgroups was considered a test dataset, while the rest were used as training and validation datasets. In the next iteration, another data subgroup was considered a test dataset, while we used the other set as a training and verification dataset. In addition, to ensure that the model starts from the optimal state at each iteration, the model\u0026rsquo;s hyperparameters were cleared after each training.\u003c/p\u003e\u003cp\u003eThe Google Colab Pro training environment was used for training CNN models. This environment contains three different GPUs, T4, T100, and A100, for our experimental models. The environment uses an NVIDIA DGX A100 with 40GB of memory and 83.5 GB of system RAM.\u003c/p\u003e\u003cp\u003e\u003cb\u003ePerformance Measurements\u003c/b\u003e\u003c/p\u003e\u003cp\u003eA confusion matrix was used to evaluate the models\u0026rsquo; performance. In our models, calculating the relevant quantity metrics for a binary classifier could be represented as a 2x2 matrix, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. This matrix\u0026rsquo;s elements are True positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN), which denotes the number of positive samples that were correctly classified, the number of negative samples that were correctly classified, the number of positive samples that incorrectly classified and the number of negative samples that incorrectly classified respectively.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFrom the confusion matrix\u0026rsquo;s elements, we used four methods to check the correctness of the model prediction. First, overall the dataset images, the ratio of correct prediction samples indicates accuracy, the accuracy in Eq.\u0026nbsp;(1). Second, using the ratio of the number of truly predicted samples for one positive class over the summation of truly and falsely the number of predicted samples for the same class. This ratio indicates recall and its Eq.\u0026nbsp;(2). Thirdly, precision, which is determined as the ratio between properly categorized samples and samples allocated to that class, indicates the percentage of the retrieved samples that are relevant as written in Eq.\u0026nbsp;(3). Finally, F1 Score is a harmonic mean of precision and recall as shown in Eq.\u0026nbsp;(4). F1 score is generally used in [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e].\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:Accuracy=\\frac{TP+TN}{TP+TN+FP+FN}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(1\\right)$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:Recall=\\frac{TP}{TP+FN}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(2\\right)$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:Precision=\\frac{TP}{TP+FP}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(3\\right)\\:\\:\\:$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:{F}_{1}=2*\\frac{Precision*Recall}{Precision+Recall}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(4\\right)\\:$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eStatistical Analysis\u003c/h2\u003e\u003cp\u003eTo enhance labeling accuracy and minimize bias, two experienced dental radiologists (E.O. and M.T.G.) independently evaluated the images. The images were re-examined using a cross-validation approach. In the Cohen\u0026rsquo;s Kappa test, this ratio was found to be approxi mately 1.0. This result shows high agreement between the two observers.\u003c/p\u003e\u003cp\u003eTen-fold cross-validation was conducted to evaluate the significance of performance differences between models, calculating mean and standard deviation for critical metrics such as F1-score, sensitivity, specificity, and accuracy. Model performance was primarily measured by accuracy, calculated as a ratio between the summation of TP and TN with the total predicted samples, and additional metrics including precision, sensitivity, and specificity. Due to the non-normal distribution of results, non-parametric tests were employed for statistical analysis.\u003c/p\u003e\u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003eThis section will cover the networks’ training results as well as our study’s majority voting effectiveness. We used several networks for tooth fracture classification and tooth detection. Regarding the detection performance, the network performed admirably. But the CNN models’ accuracy on their own wasn’t enough. We solved this by enhancing the overall accuracy using a majority vote method. We enhanced the findings by merging the predictions from many models. To highlight our results, let’s title the challenges we will solve: 1) detecting teeth in occlusal teeth images, 2) classifying detected teeth as fractured or non-fractured teeth. For the first challenge, we used the YOLOv8n model for detecting teeth. We evaluated the model by applying 20 images as a test to the model, and it detected the teeth correctly. Figure 7 contains three different images tested by the YOLOv8n model.\u003c/p\u003e\n\u003cp\u003eThe model’s performance is measured by the experimentation process’s accuracy and object detection’s robustness. In our model, the mean average precisions (mAPs) are 99.5% and 81.4% for the 50 and (50-95) thresholds, respectively. Figure 8 shows the training results for mAP50 and mAP50-95.\u003c/p\u003e\n\u003cp\u003eMany AI models have been used to train classification networks. Those models were arranged in descending order based on the accuracy of each model, as shown in Table 1. Among all the trained models, VGG19 and InceptionResNetV2 had the highest performance percentages, 87.92% and 87.44%, respectively. The rest received lower performance accuracy, as EfficientNetB0, InceptionV3, VGG16, ResNet50, ResNet50v2, and Xception obtained between 84.65% to 86.93%, which is considered a lower percentage than the previous models. For Recall, Precision, and F1-score evaluation vectors, ResNet50v2, InceptionV3, and VGG19 achieved the highest percentages, which are 92.56%, 92.02%, and 91.01%, respectively.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAs shown in Table 1, the highest accuracy needed to be higher. Therefore, we used the majority voting method to increase the prediction accuracy. While exploring the best combinations for majority voting, the accuracy of every model was considered. Firstly, we started with the best three models. After that, we moved to a combination of the highest accuracy of four models, until the best combination of seven models was identified. Finally, we used all eight models in one combination. The majority voting method was employed in each combination to determine which arrangement produced the best accuracy.\u003c/p\u003e\n\u003cp\u003eThis method of choosing the combination that gave the highest accuracy was applied consistently to all models, providing an analytical evaluation of ensemble performance in various configurations. To enhance the generality of our exploration of model combinations, we decided to use a random selection procedure to ensure the best model combination was picked out of all of our explorations. Table 2 shows the result of the combined model exploration.\u003c/p\u003e\n\u003cp\u003eAll eight models are arranged in ascending order at the beginning, shown in the first row. The subsequent rows in the table represent a distinct combination of CNN models, highlighted by light blue blocks indicating the inclusion of a particular model in that combination. Each highlighted block corresponds to the model used in the combination based on the arrangement established in the first table row. The accuracy column on the right displays the majority voting ensemble’s overall accuracy for each model combination.\u003c/p\u003e\n\u003cp\u003eStarting with the three highest model accuracies, which are VGG19, Inception-ResNetV2, and ResNet50v2, we got 90.95%. For the second combination, we added the InceptionV3 model to the previous one, and the accuracy decreased to 89.69%. The accuracy increased to 90.68% after adding the Xception model to the second combination. In the fourth combination, adding VGG16 models decreased the overall accuracy to 90.19%. However, the accuracy is reduced when the ResNet50 model is added. In the last scenario for creating majority voting combination-based models’ accuracy, the majority voting accuracy increased to 91.19% while using all models as one combination. Finally, randomly integrating VGG19, EfficientNetB0, InceptionResNetV2, InceptionV3, and Xception models yielded the highest accuracy of 91.94%.\u003c/p\u003e\n\u003cp\u003eFrom Table 2, we can conclude that the results demonstrate the effectiveness of majority voting as an ensemble method. The total accuracy is significantly increased using the combined collective decisions of many models. The rising accuracy percentages highlight the benefits of mixing several models, which improves overall performance.\u003c/p\u003e\n\u003cp\u003eTable 3 shows the fusion evaluation vectors. Table 3 shows that the highest accuracy among the five CNN models is VGG19’s accuracy, which is 87.92. In comparison with VGG19 accuracy, the Fusion accuracy is increased to 91.94%. Also, for recall, the fusion recall evaluation is 94.14% compared to the highest five models' recall, which is 92%. Moreover, for precision, the CNN model’s highest precision evaluation is 92%, and the fusion’s precision is 93.41%, which is higher than the CNN model’s precision evaluation. Finally, the fusion f1-score evaluation is 93.77%, which is higher than the highest CNN model, which is 91% for VGG19.\u003c/p\u003e\n\u003cp\u003eWe used K-fold cross-validation to evaluate the network models. Datasets are divided into 10 dataset subgroups to ensure the model’s generalization gives us a reliable and robust estimation. Figure 9 is the confusion matrix for the general model; the numbers of fractured (TP) and non-fractured (TN) samples that were predicted correctly are 241 and 15, respectively. On the other hand, the number of fractured samples, but the model predicted as non-fractured (FN), is 15; moreover, 17 samples are predicted as non-fractured and fractured (FP).\u003c/p\u003e\n\u003cp\u003eAlthough the fusion model demonstrated improved performance across all evaluation metrics, we acknowledge the limitations inherent in training on a relatively small dataset of 398 cropped images. Moreover, given the binary classification task, we prioritized recall and F1-score in our analysis, as these metrics are more clinically meaningful. In particular, maximizing recall reduces the likelihood of false negatives, which is critical in medical diagnostics to avoid missed fracture cases.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003ePatients who experience dental trauma may face various challenges in obtaining a timely and accurate diagnosis. These challenges can stem from the need for rapid post-trauma assessment, the limited availability of specialized dentists, variations in individual diagnostic experience, and the inherent limitations of radiographic imaging techniques. Although most modern dental clinics are equipped with radiographic imaging systems, the impact of these techniques on diagnostic accuracy should not be overlooked. At this point, AI-based analysis systems have the potential to assist clinicians in dental trauma assessments by standardizing the diagnostic process and enhancing accuracy.\u003c/p\u003e\u003cp\u003eVarious studies can detect fractures in jaw bones or teeth using AI models. Dong-Min Son et al. [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], offer automated mandibular fracture detection using DL techniques. This work suggests using DL methods to detect mandibular fractures from a panoramic radiograph without the need for radiologists\u0026rsquo; assistance. You Only Look Once (YOLO) is a one-stage detection technique that is the DL system utilized in this work. Panoramic radiographs were pre-processed using gamma modulation, multi-bounding boxes, single-scale luminance adaptation transform (SLAT), and multi-scale luminance adaptation transform (MLAT) to increase the detection accuracy of the system. In the training and test datasets, the fractures\u0026rsquo; form and position are not consistent in location and shape. In addition, mandibular fractures in panoramic radiographs show considerable curvature at the background level, and the original panoramic radiographs are dark. Data augmentation and preprocessing approaches mentioned above demonstrated detection accuracy for mandibular fractures with ambiguous forms and regions.\u003c/p\u003e\u003cp\u003eAnother study conducted by Son et al. [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] used combined deep-learning techniques to aid in the diagnosis of mandibular fractures. They used combined DL techniques, which are YOLO and U-Net algorithms. Without CBCT, mandibular fractures were detected using the auxiliary diagnostic algorithms YOLO and U-Net based on panoramic images. The purpose of combining these algorithms in this network is that each one has an advantage that the other does not. They used U-net because it helps segment widely dispersed fracture locations. Hence, it is utilized in conjunction with U-net to complement undiagnosed instances of mandibular fracture, while YOLO is weak in this field. At the same time, they used YOLO because of its advantage in detecting fracture locations.\u003c/p\u003e\u003cp\u003eMoreover, some studies use AI to detect fractures in the maxillofacial region. Warin et al. [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e] evaluated CT images in 4 classes: frontal fracture, midface fracture, mandibular fracture, and no fracture. Warin et al. [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e] examined midface fractures in CT images, while other studies evaluated fractures in the mandibular bone in panoramic images [\u003cspan additionalcitationids=\"CR30\" citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. In Nishiyama et al. [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e], used DL models to detect fractures in the mandibular condyle region in panoramic images. Some studies use DL models to detect tooth root fractures. Some studies can detect vertical root fractures in CBCT images [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e], or panoramic images [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] using DL models. Chang et al. [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e] used AI architectures to predict the probability of root fracture after a crown or root canal treatment. Three different CNN architectures were used in a study that detected fractures in dental implants in panoramic and periapical radiographs [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eFew studies in the literature have explored the application of AI in the domain of dental traumatology. Notably, Bani-Hani et al. [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e] introduced a CNN-based model using periapical radiographs to classify dental fractures into four subtypes. However, their work lacked automated tooth localization, relied on a relatively small dataset (108 fractured teeth), and reported a moderate overall accuracy of 78.7%. In contrast, the present study, for the first time, leverages OR for fracture detection, employs YOLOv8-based automatic localization of the maxillary central incisors, and applies an ensemble classification approach combining five CNN architectures through majority voting. With a more extensive and balanced dataset (398 teeth), significantly higher classification accuracy (91.94%), and potential for real-time application, the proposed method provides a technically and clinically superior diagnostic tool. These attributes underscore the originality and practical relevance of our approach in comparison to prior literature.\u003c/p\u003e\u003cp\u003eIn the current study, tooth fractures were detected using OR images. During the training phase of our tooth detection models, we experimented with multiple versions of the YOLO algorithms, including YOLOv8, YOLOv9, and YOLOv10, to assess their effectiveness in accurately detecting tooth localizations. Throughout the process, the training and testing accuracy metrics for these versions were consistently high and showed minimal variance, indicating that models had effectively learned the patterns within the dataset. However, while evaluating these models on test samples, YOLOv8 showed the best robustness and stability, outperforming YOLOv9 and YOLOv10. This robustness behavior in YOLOv8 leads to more generalizability in many different data scenarios, and we selected this model as our primary tooth detection approach.\u003c/p\u003e\u003cp\u003eIn the subsequent step of the proposed study, the locations of the 11 and 21 teeth, identified using the YOLOv8 detector, were utilized for classification. Contrary to the detection step, no pre-processing was employed during the classification. Various pre-trained classifier networks were employed, and accuracy rates of 84.65% and 87.92% were obtained. To further improve those results, majority voting was also utilized in the study. It allows the ensemble to compensate for the limitations of individual models, capturing a broader range of patterns and achieving better generalization. Through the application of majority voting, the overall accuracy increased by 4.6\u0026ndash;91.94%.\u003c/p\u003e\u003cp\u003eIn the present study, tooth fracture detection was performed on maxillary central incisors using the YOLOv8 algorithm, achieving an accuracy rate of 91.94%. This outcome demonstrates that image-based AI models offer more direct, precise, and clinically applicable solutions for dental trauma diagnosis compared to traditional large language models (LLMs). For instance, \u0026Ouml;zden et al. [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e] reported that ChatGPT and Google Bard (Gemini), two widely used general-purpose LLMs, provided correct responses to trauma-related dental questions with only 57.5% accuracy, a figure falling below the acceptable threshold for clinical reliability. Similarly, in a study by Tokg\u0026ouml;z Kaplan and Cankar [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e] focusing on dental avulsion, the Gemini model outperformed ChatGPT-3.5 in terms of overall accuracy (p\u0026thinsp;\u0026lt;\u0026thinsp;0.004); however, both models exhibited limited success in responding to multiple-choice technical questions. These findings highlight the inherent limitations of LLMs in clinical decision-making, particularly in domains that require visual interpretation and domain-specific training. In contrast, our use of the YOLOv8 model, trained directly on orthopantomographic (OPG) images, allowed for accurate localization and classification of dental fractures, demonstrating the model\u0026rsquo;s practical utility as a decision-support tool. Furthermore, the scope of our study extends beyond model architecture to include expert-driven data curation, annotation, and the application of ensemble learning (majority voting) to enhance classification robustness. In this regard, our work distinguishes itself methodologically and technically from prior LLM-based approaches and underscores the potential of image-based deep learning systems to provide more effective and reliable diagnostic support in the field of dental traumatology.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eAccurate detection and classification of dental fractures are crucial for effective diagnosis and treatment planning. In this study, an AI-based automated system was developed for detecting fractures in maxillary central incisors. The proposed method achieved an accuracy rate of 91.94% by utilizing YOLOv8 for tooth detection and CNN-based models for fracture classification. This study is significant as one of the first investigations employing DL techniques for dental fracture detection using OR. The developed AI-based automated approach can assist clinicians in making faster and more consistent diagnoses, reducing interobserver variability. The integration of AI-assisted tools in dentistry holds the potential to enhance diagnostic accuracy and improve patient outcomes.\u003c/p\u003e\u003cp\u003e\u003cb\u003eLimitations\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThis study has several limitations. First, only OR was utilized, without validation through other imaging techniques such as periapical radiography or CBCT. As a result, the three-dimensional structure and orientation of fractures could not be fully assessed. Additionally, the dataset was obtained from a single center, limiting the generalizability of the model to different patient populations. Moreover, the study exclusively focused on crown fractures, classifying them as either \u0026ldquo;fractured\u0026rdquo; or \u0026ldquo;non-fractured,\u0026rdquo; without considering root fractures or other dental anomalies. Finally, the model was trained and tested solely on maxillary central incisors (11 and 21), and its performance on other tooth groups was not evaluated.\u003c/p\u003e\u003cp\u003eConsidering these limitations, further improvements in detection and classification accuracy are required for the proposed method to be effectively integrated into clinical practice. Future studies should incorporate additional imaging modalities, such as periapical radiography and CBCT, to enhance the assessment of fracture orientation and depth. Furthermore, evaluating the model using larger and more diverse datasets from multiple centers will be crucial for improving its generalizability. Expanding the classification to include both crown and root fractures, along with their subtypes, could provide a more comprehensive diagnostic framework. Additionally, testing the model on different tooth groups would further validate its applicability and diagnostic performance.\u003c/p\u003e\u003cp\u003eThe originality of this work stems from the problem itself and the comprehensive analysis applied to it, rather than from architectural innovation. Substantial effort was dedicated to the collection, curation, and meticulous labeling of the dataset, which required considerable domain expertise and time. Given the resource-intensive nature of this process, we made a deliberate decision to focus on fine-tuning well-established, high-performing architectures rather than developing new ones, in order to ensure that our results would be robust, reproducible, and directly comparable to the state-of-the-art.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAhmed N. Nusari: Writing – review \u0026amp; editing, Writing – original draft, Methodology, Investigation, Conceptualization. Esra ÖNCÜ: Writing – review \u0026amp; editing, Supervision, Investigation, Conceptualization. Emin Argun Oral: Writing – original draft, Methodology, Investigation, Conceptualization. İbrahim Yücel Özbek: Writing – original draft, Investigation, Conceptualization. Özkan MİLOĞLU: Writing – original draft, Final approval. Mustafa Taha GÜLLER: Final approval.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eThis research received no external funding.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical approval\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eEthics committee approval was obtained from Atatürk University Faculty of Dentistry Ethics Committee, date: 2023/ number: 61.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration of competing interest\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data that support the findings of this study are available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInformed consent\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cu\u003eClinical trial number\u003c/u\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cu\u003eNot applicable\u003c/u\u003e\u003c/strong\u003e\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBourguignon C, Cohenca N, Lauridsen E, Flores MT, O'Connell AC, Day PF, et al. International Association of Dental Traumatology guidelines for the management of traumatic dental injuries: 1. Fractures and luxations. Dent Traumatol. 2020;36(4):314\u0026ndash;30.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJogezai U, Kalsi A. Long-term complications and management of dental trauma in the adult patient-Part 1: fractured teeth, pulpal complications and resorption. Br Dent J. 2024;237(2):95\u0026ndash;105.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKullman L, Al Sane M. Guidelines for dental radiography immediately after a dento-alveolar trauma, a systematic literature review. Dent Traumatol. 2012;28(3):193\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePatel S, Puri T, Mannocci F, Navai A. Diagnosis and management of traumatic dental injuries using intraoral radiography and cone-beam computed tomography: an in vivo investigation. J Endod. 2021;47(6):914\u0026ndash;23.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHuang C, Wang J, Wang S, Zhang Y. A review of deep learning in dentistry. Neurocomputing. 2023;554:126629.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHu Z, Cao D, Hu Y, Wang B, Zhang Y, Tang R, et al. Diagnosis of in vivo vertical root fracture using deep learning on cone-beam CT images. BMC Oral Health. 2022;22(1):382\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAbdelazim R, Fouad EM. Artificial intelligent-driven decision-making for automating root fracture detection in periapical radiographs. BDJ Open. 2024;10(1):76.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSarkar M, Mandal A. Slaahe: selective apex adaptive histogram equalization. Frankl Open. 2023;3:100023.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTerven J, C\u0026oacute;rdova-Esparza DM, Romero-Gonz\u0026aacute;lez JA. A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo-nas. Mach Learn Knowl Extr. 2023;5(4):1680\u0026ndash;716.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSary IP, Andromeda S, Armin EU. Performance comparison of YOLOv5 and YOLOv8 architectures in human detection using aerial images. Ultima Computing: J Syst Comput. 2023;15(1):8\u0026ndash;13.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePaymode AS, Malode VB. Transfer learning for multi-crop leaf disease image classification using convolutional neural network VGG. Artif Intell Agric. 2022;6:23\u0026ndash;33.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu Y, Wang J. Transfer learning based multi-layer extreme learning machine for probabilistic wind power forecasting. Appl Energy. 2022;312:118729.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRoy AM. Adaptive transfer learning-based multiscale feature fused deep convolutional neural network for EEG MI multiclassification in brain\u0026ndash;computer interface. Eng Appl Artif Intell. 2022;116:105347.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVignesh S, Savithadevi M, Sridevi M, Sridhar R. A novel facial emotion recognition model using segmentation VGG-19 architecture. Int J Inform Technol. 2023;15(4):1777\u0026ndash;87.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMohamed ISE, Zarif S, Abdel-Kader H, Elsayed M. Breast cancer detection based on hybrid deep learning models. IJCI Int J Computers Inform. 2023;10(3):119\u0026ndash;25.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBhardwaj P, Kumar S, Kumar Y. A comprehensive analysis of deep learning-based approaches for the prediction of gastrointestinal diseases using multi-class endoscopy images. Arch Comput Methods Eng. 2023;30(7):4499\u0026ndash;516.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDas B, Saha A, Mukhopadhyay S. Rain removal from a single image using refined inception ResNet v2. Circuits Syst Signal Process. 2023;42(6):3485\u0026ndash;508.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSrinivas K, Gagana Sri R, Pravallika K, Nishitha K, Polamuri SR. COVID-19 prediction based on hybrid Inception V3 with VGG16 using chest X-ray images. Multimedia Tools Appl. 2024;83(12):36665\u0026ndash;82.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu Y, Zhang L, Hao Z, Yang Z, Wang S, Zhou X, et al. An xception model based on residual attention mechanism for the classification of benign and malignant gastric ulcers. Sci Rep. 2022;12(1):15365.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJhang K. Voting and ensemble schemes based on CNN models for photo-based gender prediction. J Inform Process Syst. 2020;16(4):809\u0026ndash;19.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJain A, Kumar A, Susan S. Evaluating deep neural network ensembles by majority voting cum meta-learning scheme. Springer. 2022;29\u0026ndash;37.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBerrar D, Cross-Validation. Encyclopedia Bioinf Comput Biology. 2019;1:542\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAlaoui SS, Labsiv Y, Aksasse B. Classification algorithms in data mining. Int J Tomography Simul. 2018;31(4):34\u0026ndash;44.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHicks SA, Str\u0026uuml;mke I, Thambawita V, Hammou M, Riegler MA, Halvorsen P, et al. On evaluation metrics for medical applications of artificial intelligence. Sci Rep. 2022;12(1):5979.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSon DM, Yoon YA, Kwon HJ, An CH, Lee SH. Automatic detection of mandibular fractures in panoramic radiographs using deep learning. Diagnostics. 2021;11(6):933.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSon DM, Yoon YA, Kwon HJ, Lee SH. Combined deep learning techniques for mandibular fracture diagnosis assistance. Life. 2022;12(11):1711.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWarin K, Limprasert W, Suebnukarn S, Paipongna T, Jantana P, Vicharueang S. Maxillofacial fracture detection and classification in computed tomography images using convolutional neural network-based models. Sci Rep. 2023;13(1):3434.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWarin K, Vicharueang S, Jantana P, Limprasert W, Thanathornwong B, Suebnukarn S. Deep learning for midfacial fracture detection in CT images. MEDINFO 2023-The Future Is Accessible. IOS; 2024. pp. 1497\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVinayahalingam S, van Nistelrooij N, van Ginneken B, Bressem K, Tr\u0026ouml;ltzsch D, Heiland M, et al. Detection of mandibular fractures on panoramic radiographs using deep learning. Sci Rep. 2022;12(1):19596.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYari A, Fasih P, Hosseini Hooshiar M, Goodarzi A, Fattahi SF. Detection and classification of mandibular fractures in panoramic radiography using artificial intelligence. Dentomaxillofacial Radiol. 2024;53(6):363\u0026ndash;71.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWarin K, Limprasert W, Suebnukarn S, Inglam S, Jantana P, Vicharueang S. Assessment of deep convolutional neural network models for mandibular fracture detection in panoramic radiographs. Int J Oral Maxillofac Surg. 2022;51(11):1488\u0026ndash;94.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNishiyama M, Ishibashi K, Ariji Y, Fukuda M, Nishiyama W, Umemura M, et al. Performance of deep learning models constructed using panoramic radiographs from two hospitals to diagnose fractures of the mandibular condyle. Dentomaxillofacial Radiol. 2021;50(7):20200611.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYang P, Guo X, Mu C, Qi S, Li G. Detection of vertical root fractures by cone-beam computed tomography based on deep learning. Dentomaxillofacial Radiol. 2023;52(3):20220345.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFukuda M, Inamoto K, Shibata N, Ariji Y, Yanashita Y, Kutsuna S, et al. Evaluation of an artificial intelligence system for detecting vertical root fracture on panoramic radiography. Oral Radiol. 2020;36:337\u0026ndash;43.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChang WT, Huang HY, Lee TM, Sung TY, Yang CH, Kuo YM. Predicting root fracture after root canal treatment and crown installation using deep learning. J Dent Sci. 2024;19(1):587\u0026ndash;93.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLee DW, Kim SY, Jeong SN, Lee JH. Artificial intelligence in fractured dental implant detection and classification: evaluation using dataset from two dental hospitals. Diagnostics. 2021;11(2):233.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBani-Hani T, Wedyan M, Al-Fodeh R, Shuqeir R, Al Jundi S, Tewari N. Artificial intelligence model for application in dental traumatology. Eur Arch Paediatr Dent. 2025 May;31. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s40368-025-01063-0\u003c/span\u003e\u003cspan address=\"10.1007/s40368-025-01063-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOzden I, Gokyar M, Ozden ME, Sazak Ovecoglu H. Assessment of artificial intelligence applications in responding to dental trauma. Dent Traumatol. 2024;40(6):722\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTokg\u0026ouml;z Kaplan T, Cankar M. Evidence-based potential of generative artificial intelligence large language models on dental avulsion: ChatGPT versus Gemini. Dent Traumatol. 2025;41(2):178\u0026ndash;86.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTables 1 to 3 are available in the Supplementary Files section.\u003c/p\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":"bmc-oral-health","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ohea","sideBox":"Learn more about [BMC Oral Health](http://bmcoralhealth.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/ohea/default.aspx","title":"BMC Oral Health","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Dental, Tooth Fracture, Tooth Detection, CNN models, YOLOv8, Occlusal Radiography, Majority Voting","lastPublishedDoi":"10.21203/rs.3.rs-7207126/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7207126/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cb\u003eBackground\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThis study highlights the importance of accurate and quick diagnosis of detection and classification of fractured teeth and the potential benefits of applying deep learning (DL) techniques to solve this problem.\u003c/p\u003e\u003cp\u003e\u003cb\u003eMethods\u003c/b\u003e\u003c/p\u003e\u003cp\u003eIn the study, occlusal radiography (OR) of the premaxilla is used for tooth fracture detection of teeth numbers 11 and 21. For that, a dataset that contains 200 ORs of various tooth conditions was constructed. In the proposed method, teeth numbers 11 and 21 are automatically detected in OR images using a YOLOv8-based machine learning framework as the first step. Then, images of these two teeth are obtained by cropping OR images using the bounding box coordinates of numbers 11 and 21 teeth, obtained by a YOLO-based detector. Finally, these cropped images are provided as input to a pre-trained CNN-based network to classify between the \u0026ldquo;fracture\u0026rdquo; or \u0026ldquo;non-fractured tooth\u0026rdquo;. For this purpose, VGG19, EfficientNetB0, InceptionResNetV2, and InceptionV3 nets are employed, and the obtained classification results are fused by applying a majority voting step to further improve the performance.\u003c/p\u003e\u003cp\u003e\u003cb\u003eResults\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe experimental studies obtained a 99.5% mean average precision (mAP50) score. On the other hand, percent accuracy rates in the range from 84.65 to 87.92 were observed using five pre-trained networks, and the percent accuracy metric was improved to 91.94% using the majority voting-based fusion approach.\u003c/p\u003e\u003cp\u003e\u003cb\u003eConclusion\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe findings indicate that the proposed method effectively detects fractured teeth by leveraging machine learning techniques. Furthermore, this approach provides a pioneering framework for integrating artificial intelligence (AI) methodologies into dental diagnostics, offering clinicians a reliable decision-support tool for improved diagnostic accuracy.\u003c/p\u003e","manuscriptTitle":"Deep Learning-Based Fractured Tooth Detection in Occlusal Radiographs","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-03 01:11:09","doi":"10.21203/rs.3.rs-7207126/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-10-10T09:48:28+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-10-07T12:09:10+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"150684206336167363014912328746148694294","date":"2025-10-06T11:25:13+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-10-06T05:51:02+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"261537777813060949266857850225801608170","date":"2025-10-06T05:42:16+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"25440854858833695066404229052825206500","date":"2025-10-06T02:50:02+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-09-30T06:34:48+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"85408637083457075927079916677478120796","date":"2025-09-26T13:45:00+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"81829872719889701585002411590362999865","date":"2025-09-22T04:09:26+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-09-21T20:12:23+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-09-16T08:35:11+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-08-04T03:37:39+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-08-04T03:37:26+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Oral Health","date":"2025-07-24T15:32:27+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-oral-health","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ohea","sideBox":"Learn more about [BMC Oral Health](http://bmcoralhealth.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/ohea/default.aspx","title":"BMC Oral Health","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e4202168-d829-4e85-8736-8ddbe1bf1e3a","owner":[],"postedDate":"October 3rd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-12-22T15:59:42+00:00","versionOfRecord":{"articleIdentity":"rs-7207126","link":"https://doi.org/10.1186/s12903-025-07445-6","journal":{"identity":"bmc-oral-health","isVorOnly":false,"title":"BMC Oral Health"},"publishedOn":"2025-12-18 15:57:12","publishedOnDateReadable":"December 18th, 2025"},"versionCreatedAt":"2025-10-03 01:11:09","video":"","vorDoi":"10.1186/s12903-025-07445-6","vorDoiUrl":"https://doi.org/10.1186/s12903-025-07445-6","workflowStages":[]},"version":"v1","identity":"rs-7207126","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7207126","identity":"rs-7207126","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}