Two-stage deep learning approach for screening disk displacement of the temporomandibular joint using orthopantomogram

Two-stage deep learning approach for screening disk displacement of the temporomandibular joint using orthopantomogram | 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 Article Two-stage deep learning approach for screening disk displacement of the temporomandibular joint using orthopantomogram Jong-Soo Kim, Taehyeong Kim, Yong-Hee Ahn, Hong-Gee Kim, Sug-Joon Ahn This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7053119/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 07 Jan, 2026 Read the published version in Scientific Reports → Version 1 posted 14 You are reading this latest preprint version Abstract This study aimed to develop a deep learning-based approach for the initial screening of disk displacement (DD) of the temporomandibular joint (TMJ) using orthopantomograms (OTG). A two-stage deep learning model was proposed: first, regions of interest were detected from 1,469 orthopantomogram images using YOLOv5s; second, DD status was identified from 1,564 joints with the magnetic resonance imaging-verified DD status using ResNet-18 and DANet. Diagnostic performance was evaluated through multi-class and binary classification analyses. In three-class classification (normal disk position, DD with reduction, and DD without reduction), the model achieved an overall accuracy of 71.86%. It performed well in identifying normal disk position (F1 score: 78.57%) and DD without reduction (F1 score: 80.99%) but showed lower performance in detecting DD with reduction (F1 score: 43.84%). In the binary classification, where DD with reduction and DD without reduction were combined into a single class, the model demonstrated improved accuracy (82.04%), sensitivity (78.85%), and specificity (87.30%). The model exhibited potential in estimating DD severity (R² = 0.4871). Given its strong ability to differentiate between normal disk position and DD in a binary classification setting, this tool has the potential to serve as an initial screening tool for TMJ DD in clinical practice. Biological sciences/Computational biology and bioinformatics Health sciences/Diseases Health sciences/Health care Health sciences/Medical research Temporomandibular joint Disk displacement Screening Orthopantomogram Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction The temporomandibular joint (TMJ) is configured to withstand repetitive and high functional loads between the temporal bone and the mandibular condyle. The articular disk, positioned between the condyle and temporal bone, serves as a stress distributor during jaw activity 1 . Consequently, disk displacement (DD) can increase stress on surrounding tissues, leading to damage to the articulating surfaces of the condyle and temporal bone 2 . Therefore, DD is associated with degenerative changes in TMJ articulating structures, manifesting as various symptoms such as joint noise, pain, and restricted mouth opening 2 , 3 . In addition, TMJ DD influences aspects of the stomatognathic system, such as condylar and mandibular growth 4 , recovery and regeneration of the condyle 5 , and centric relation-maximum intercuspation discrepancy 6 , which may have a significant effect on the results of oral rehabilitation 7 , 8 . Therefore, it is important to screen for TMJ DD prior to commencing oral rehabilitation. Magnetic resonance imaging (MRI) is the gold standard for determining the position of the articular disk due to its superior ability to depict the disk 9 . However, its clinical use in dentistry is limited by cost-effectiveness concerns, spatial constraints, and various contraindications 10 . While diagnostic criteria for temporomandibular disorders based on clinical variables provide an alternative, they require time-consuming training and examinations 11 . As a result, there is a strong need for simpler and more cost-effective methods to screen for TMJ DD. A strong correlation between DD and degenerative condylar changes has been reported using MRI and computed tomography 12 – 14 . Condyles affected by DD exhibit a higher prevalence of degenerative changes, such as erosion, osteophyte formation, and flattening, compared to those with normal disk positioning (NR) 14 . These changes are associated with altered condylar size and shape, and joints with DD have reduced condylar width, height, length, and volume compared to those with NR 12 , 13 . Orthopantomograms (OTGs) are widely used for the initial screening of condylar pathologies due to their low cost, minimal radiation exposure, and ability to provide a comprehensive view of the joint structure 15 . Given that TMJ DD is associated with degenerative condylar changes and altered condylar morphologies 12 – 14 , OTG may serve as a viable tool for detecting DD-related structural changes of the mandibular condyle. A previous study also reported that specific condylar morphologies, such as decreased condylar height and a distally inclined condyle, were frequently observed in OTGs of patients with TMJ DD 16 . Recently, deep learning approaches have been increasingly utilized in dentistry to improve the diagnosis of complex oral diseases 17 , 18 . While TMJ DD screening models have been developed using MRI or clinical data 2 , 19 , the potential of deep learning applied to OTG images remains unexplored. The purpose of this study was 1) to develop a deep learning-based approach for screening TMJ DD using OTG and 2) to evaluate the performance of this approach. Methods Study design We analyzed data from patients who visited Seoul National University Dental Hospital between 1994 and 2023. The inclusion criteria were as follows: 1) undergoing OTG for a general examination, 2) no history of TMJ trauma, TMJ treatment, orthodontic treatment, or orthognathic surgery, and 3) no systemic disease, including rheumatoid arthritis and congenital anomalies. This retrospective study was approved by the Institutional Review Board of the University (S-D20220026). The requirement for informed consent was waived due to the retrospective nature of this study using existing records. This study was conducted in accordance with the Declaration of Helsinki. We proposed a two-stage deep learning approach, TMJ-PanoNet, for identifying TMJ DD using OTG. This approach utilized two distinct datasets: one for training a model to detect regions of interest (ROI) and another for TMJ DD identification. The ROI detection dataset included 1,469 OTG images from 598 males and 871 females who underwent only OTG. The TMJ DD identification dataset comprised 1,564 joints from 229 males and 570 females who underwent both OTG and TMJ MRI within 3 months. For the DD identification dataset, participants were limited to those aged 15–50 years, as active condylar growth during puberty can influence condylar size and shape 19 , and participants over 50 years were excluded due to a small sample size. The inclusion and exclusion process for screening for TMJ DD is shown in a flowchart in Fig. 1 . OTG image acquisition OTG images were acquired using the OP-100 (Instrumentarium, Tuusula, Finland). Patients’ heads were positioned optimally using standard head positioning devices and bite blocks, following the manufacturer’s instructions. Imaging parameters were set to 70 kVp, 10 mA, and an exposure time of 17.6 s. OTG data were stored in TIF format, either by scanning analogue images or downloading digital images from a web-based picture archiving and communication system. Disk position determination TMJ MRI was performed using Signa Horizon (GE, Waukesha, WI) or Magnetom Vita (Siemens, Erlangen, Germany) ( Supplementary Table S1 online ). Reference annotations for DD categories were provided by two radiologists with over 20 years of experience in TMJ MRI interpretation, who independently evaluated disk positions in a blinded manner. Following a thorough assessment, DD was classified into three categories: normal disk position (NR), DD with reduction (DDR), and DD without reduction (DDNR) as previously described 13 , 16 . Any discrepancies in disk position classification were resolved by consensus. Condyles with partial DD, DD with partial reduction, pure-sideway DD, and posterior DD were excluded due to their small sample sizes. For continuous severity assessment, DD severity scores were assigned based on clinical progression: NR was assigned a score of 0 (no displacement), DDR a score of 1 (intermediate severity), and DDNR a score of 2 (most severe condition). Determination of ROI in OTG Images We developed a specialized annotation tool for precise identification of anatomical landmarks to facilitate accurate detection of ROI around the mandibular condyle in OTG images (Fig. 2 a). The tool’s interface included options to navigate through images, clear points, and save annotations, facilitating efficient and consistent landmark identification across multiple images. Five anatomical landmarks were identified in each region (Fig. 2 b) to define the ROI around the condylar area associated with DD (Figs. 2 c-h). An experienced orthodontist with over 25 years of practice carefully marked these landmarks, which directly established the ROI boundaries across all images. The annotated regions then served as ground truth data for training the ROI detection model in the first stage of TMJ-PanoNet. Dataset partitioning for model development Previous studies have shown that the relationship between condylar morphology and DD is not significantly influenced by sex 13 , 14 ; therefore, data from both males and females were combined. For model training and evaluation, both datasets were divided into training, validation, and testing subsets. The ROI detection dataset of 1,469 OTG images was randomly split into training (1,028 images), validation (293 images), and testing (148 images) subsets. The DD classification dataset was also divided into training (1,244 joints), validation (153 joints), and testing (167 joints) subsets. These two datasets were completely separate with no patient overlapping. Additionally, for the DD classification dataset, patient-level separation was implemented to prevent data leakage, ensuring that both condyles from the same patient were assigned to the same subset. This rigorous segregation ensured that the testing dataset consisted of completely unseen patients, reflecting model performance on new cases and mimicking real clinical scenarios. Architecture of TMJ-PanoNet TMJ-PanoNet was designed as a two-stage deep learning model (Fig. 3 a). The first stage focused on the ROI detection, whereas the second stage analyzed these ROIs to classify DD status. In the first stage, YOLOv5s 20 was trained to automatically detect and localize the ROIs for both the left and right sides of the TMJ using the annotated regions in OTG images. This approach enabled effective localization across patients with varying mandibular structures, regardless of the variations in scales and orientations of the mandibular condyles. Before passing the ROIs to the second stage for TMJ DD classification, the left condylar region was horizontally flipped to match the orientation of the right condylar region. Additionally, all detected ROIs were resized to a fixed resolution ( e.g. , 224×224 pixels). This standardization ensured uniform orientation across samples and normalized condylar features for consistent analysis. Furthermore, this approach improved data homogeneity by treating left and right condyles identically, which helped address the challenge of limited data availability and enhanced the model’s generalization capability. The second stage, operating on these detected ROIs, performed classification of TMJ DD using a pretrained ResNet-18 21 for feature extraction, which was fine-tuned on our dataset, combined with position and channel attention modules of DANet 22 to emphasize crucial features. ResNet-18 was chosen due to its balance between efficiency and performance, with fewer parameters compared to deeper variants to mitigate overfitting risk, while maintaining effective feature learning through its residual connections. The DANet’s position and channel attention modules enhanced the identification of crucial anatomical features in the condylar regions (Figs. 2 c-h). For model training, we utilized a hybrid loss function approach by combining mean-variance loss 23 with standard cross-entropy loss. The feature maps were then processed through fully-connected layers (256 and 128 neurons) with batch normalization and dropout (p = 0.5) for regularization. Detailed model training procedures and implementation details are provided in the Supplementary Materials. Two-stage inference process TMJ-PanoNet performed inference in two sequential stages. First, the model automatically extracted and standardized ROIs from OTG, creating uniform condylar regions despite variations in TMJ scale. The standardized orientation of condylar regions enabled independent assessment of left and right joints while maintaining consistent feature analysis. The classification model then processed these standardized inputs to estimate TMJ DD severity scores. The severity score for each sample \(\:i\) was computed as the weighted average of class labels using predicted probabilities: where \(\:K=3\) is the total number of classes, \(\:j\) represents the class labels (NR = 0, DDR = 1, DDNR = 2), and \(\:{p}_{i,j}\) denotes the predicted probability that sample \(\:i\) belongs to class \(\:j\) . These continuous scores (ranging from 0 to 2) were converted to categorical classifications using thresholds: NR (0-0.5), DDR (0.5–1.5), and DDNR (1.5-2). This method enabled both continuous severity assessment and categorical classification for comprehensive evaluation of TMJ DD status. Model evaluation The performance of TMJ-PanoNet was evaluated using an independent test set of 167 joints, representing approximately 10.7% of the total 1,564 joints in the second dataset. We conducted multi-class (NR, DDR, and DDNR) and binary classification (NR and DD) analyses, alongside DD severity classification, to assess the model's classification capabilities (Fig. 3 b). Model performance metrics included precision, recall, F1 score, and accuracy for multi-class evaluation, and sensitivity, specificity, and accuracy for binary classification. For continuous severity assessment, the coefficient of determination (R²), Pearson correlation coefficient (r), Mean Squared Error (MSE), and Mean Absolute Error (MAE) were calculated (detailed information is available in the Supplementary Material). All metrics were computed using Python version 3.8.10 with scikit-learn library version 1.2.0. Statistical analysis was performed by an expert with more than 15 years of experience in machine learning and image processing. Results Patient characteristics of this experiment The inclusion and exclusion process for TMJ DD screening is shown in a flowchart in Fig. 1 . After applying the exclusion criteria, the final ROI detection dataset included 1,469 images (871 females and 598 males; average age 14.52 ± 7.78 years). For the TMJ DD classification dataset, 1,564 joints from 799 patients (570 females and 229 males; average age 23.98 ± 5.94 years) were included. Among these, 565 joints (36.1%) had NR, 422 joints (27.0%) had DDR, and 577 joints (36.9%) had DDNR ( Table 1 ). When we attempted to train and evaluate separate models based on sex and age groups, these approaches did not result in significant improvements (data not shown). This may be because condylar changes associated with DD are not significantly influenced by sex-related factors after active growth 12-14 . As a result, male and female condyles aged 15–50 years were combined to increase data volume and diversity, thereby enhancing the model. Diagnostic performance of ROI detection The ROI detection module achieved perfect detection performance with a 100% success rate (intersection over union (IoU) ≥ 0.5 across all 293 test images), localizing condylar regions in every case. This high performance can be attributed to the relatively straightforward nature of the ROI detection task, as mandibular condyles have consistent anatomical locations within OTG images. Based on these accurately detected ROIs, the diagnostic performance for DD classification was evaluated. Diagnostic performance of multi-class classification Fig. 4a presents the confusion matrix for the three-class classification (NR, DDR, and DDNR). In the multi-class analysis ( Table 2 ), our model achieved an overall accuracy of 71.86% (120/167), with varying performance across categories. Particularly, the model demonstrated strong performance in identifying NR (precision: 71.43%, recall: 87.30%, and F1-score: 78.57%) and DDNR cases (precision: 83.05%, recall: 79.03%, and F1-score: 80.99%), indicating reliable detection of both normal and the most severe conditions. However, it showed relatively lower performance for DDR cases (precision: 51.62%, recall: 38.10%, F1-score: 43.84%), suggesting challenges in identifying this intermediate condition. The model's strength in discriminating between extreme cases was particularly evident, with misclassifications between NR and DDNR occurring in only six cases (3.59%) of the total elements in the confusion matrix elements. Diagnostic performance of binary classification To evaluate the model's performance in a clinical screening context, we conducted a binary classification analysis, where DDR and DDNR were combined into a single DD class ( Table 3 ). This simplified classification showed improved performance with an accuracy of 82.04% (137/167). The model correctly identified 55 of 63 NR cases and 82 of 104 DD cases, achieving a sensitivity of 78.85% (82/104) and a specificity of 87.30% (55/63). These results suggest that the model could be particularly useful as an initial screening for detecting the presence of DD. Diagnostic performance of continuous severity estimation Beyond categorical classification, we assessed the model's ability to estimate continuous TMJ DD severity ( Fig. 4b ), which could provide more nuanced information about disease progression. The analysis yielded an R² of 0.4871, demonstrating the model's moderate predictive power for continuous severity estimation. The Pearson r of 0.7531 demonstrated a strong positive correlation between predicted and actual values, suggesting consistent alignment between the model's estimates and clinical assessments. The model achieved an MSE of 0.3839 and an MAE of 0.3272, with these error metrics indicating reasonable prediction accuracy considering the subjective nature of severity assessment. This continuous severity estimation capability could be particularly valuable for monitoring disease progression over time. Discussion Significant efforts have been made to develop a simple and convenient tool for screening TMJ DD to aid in proper treatment planning for patients with potential TMJ disorders 2 , 3 , 9 , 11 , 24 – 27 . While MRI remains the gold standard for diagnosing TMJ DD 2 , 3 , its clinical use in dentistry is limited due to various constraints 10 . Previous studies have demonstrated that DD is frequently accompanied by degenerative changes in the mandibular condyle 2 , 19 , resulting in alterations in its size and shape 13 , 14 . Given that OTG is a simple and cost-effective imaging tool capable of providing a comprehensive view of the condyle and its surrounding structures, it may hold potential for detecting joint morphology changes associated with DD. To the best of our knowledge, this is the first study to identify TMJ DD using OTG combined with deep learning. TMJ-PanoNet classified TMJ DD into three categories—NR, DDR, and DDNR—with an overall accuracy of 71.86% (120/167) (Table 2 ). However, its ability to identify DDR was lower than its performance in identifying NR and DDNR (Table 2 ). This discrepancy may be attributed to differences in condylar morphology among the three DD categories. In DDR, the displaced disk returns to its normal position during mouth opening, restoring disk function. As a result, condylar changes associated with DD are less pronounced in DDR than in DDNR 12 , 13 , 25 . In this study, differences in condylar morphology were less distinct between NR and DDR than between NR and DDNR or DDR and DDNR, regardless of sex (Figs. 3 C-H). Since OTG has inherent limitations in detecting minor morphological changes, subtle differences between NR (Figs. 3 C-D) and DDR (Figs. 3 E-F) may be challenging to identify in the TMJ-PanoNet. Nonetheless, some morphological changes were observed in DDR cases, despite their lower severity. Specifically, the condylar width is initially affected by the resorption of the lateral pole before more extensive morphological changes occur as DD progresses 13 , 24 . Since OTG primarily captures the lateral and medial thirds of the condylar head 16 , early changes in condylar width at the lateral pole may aid in DDR identification, which explains TMJ-PanoNet's ability to detect DDR despite its lower F1 score (Table 2 ). In contrast to DDR, DDNR cases do not exhibit disk reduction upon mouth opening, resulting in total loss of disk function during jaw function. This loss is accompanied by more pronounced degenerative changes in the condylar surfaces. Consequently, radiographically detectable degenerative changes are more evident in DDNR than in DDR 14 , with DDNR cases exhibiting the smallest condyles and the most severe morphological alterations 13 . In this study, significant differences in size and shape were observed between NR (Figs. 3 C-D) and DDNR (Figs. 3 G-H), which may explain the minimal misclassification between these extreme categories (only six cases, 3.59% of the total). Given that DDNR cases have a higher likelihood of developing osteoarthritis compared to NR and DDR cases 27 , TMJ-PanoNet’s superior ability to detect DDNR may assist clinicians in diagnosing and planning treatments for patients with TMJ DD. When DDR and DDNR were treated as a single category, the model achieved improved classification performance (Table 3 ). Specifically, the accuracy of binary classification (82.04%) was approximately 10% higher than that of the three-class classification (71.86%). The model’s accuracy, sensitivity, and specificity are around 80%, suggesting that TMJ-PanoNet may be useful for the initial screening of TMJ DD in clinical settings. Beyond condylar changes, the morphology of the articular eminence and glenoid fossa is also significantly influenced by TMJ DD. A flattened articular eminence and a larger glenoid fossa have been associated with TMJ DD 28 . In this study, the posterior slope of the articular eminence and the glenoid fossa were included in the ROI (Figs. 3 C-H), potentially contributing to the model’s improved performance in identifying DD using OTG. Additionally, TMJ-PanoNet demonstrated potential for estimating TMJ DD severity on a continuous scale (Fig. 4 b). The model achieved an R² of 0.4871, a Pearson correlation coefficient of 0.7531, an MSE of 0.3839, and an MAE of 0.3272. These results indicate a moderate to strong correlation between predicted and actual values, highlighting the model’s ability to capture subtle differences in DD severity with reasonable accuracy. This study has some limitations. First, our dataset was derived from a single OTG device, restricting data variability. To ensure the model's generalizability, future studies should incorporate data from multiple OTG devices. However, previous research has shown that standardized head positioning and bite blocks yield acceptable reproducibility of angular and linear dimensions, regardless of the OTG device used 29 . Moreover, condylar morphology associated with TMJ DD is not dependent on OTG machines 16 , 30 , suggesting that TMJ-PanoNet’s predictive performance may not be significantly influenced by device variations. Second, our study population was limited to individuals aged 15 to 50 years. While this age group represents a significant portion of patients with TMJ DD in dental clinics, further studies including a broader age range and data from multiple OTG devices are necessary to enhance the model’s clinical applicability. TMJ-PanoNet achieved a diagnostic accuracy of 71.86% in the three-class classification (NR, DDR, and DDNR). However, its performance significantly improved, reaching 82.04% accuracy, when DDR and DDNR were treated as a single category. Given its strong ability to differentiate between NR and DD in a binary classification setting, TMJ-PanoNet has the potential to serve as an initial screening tool for TMJ DD in clinical practice. The model may be particularly beneficial for clinicians who are not familiar with TMJ MRI, providing a simple, fast, and convenient decision-support system. Declarations Funding This work was supported by the New Faculty Startup Fund from Seoul National University and Seoul National University Dental Hospital (02-2025-0508). Author contributions Conceptualization: S.J.A., T.K., H.G.K.. Writing—original draft: J.S.K., T.K., S.J.A.. Writing—revision, review & editing: J.S.K., T.K., Y.H.A., H.G.K., S.J.A. Formal analysis: J.S.K., T.K. Funding acquisition: T.K., S.J.A. Investigation: J.S.K., T.K. Supervision: Y.H.A., H.G.K., S.J.A. Validation: T.K., Y.H.A., H.G.K., S.J.A. Project administration: T.K., H.G.K., S.J.A. All authors reviewed the manuscript. Data availability statement The data that support the findings of this study are available from the corresponding author upon reasonable request. Competing interests The authors declare that they have no competing interests. References Kuroda, S. et al. Biomechanical and biochemical characteristics of the mandibular condylar cartilage. Osteoarthritis. Cartilage. 17 , 1408-1415 (2009). Lee, Y.-H., Won, J. H., Kim, S., Auh, Q.-S. & Noh, Y.-K. Advantages of deep learning with convolutional neural network in detecting disc displacement of the temporomandibular joint in magnetic resonance imaging. Sci. Rep. 12 , 11352 (2022). Maracci, L. M. et al. Influence of examiner calibration on clinical and mri diagnosis of temporomandibular joint disc displacement: a systematic review and meta-analysis. Dentomaxillofac. Radiol. 53 , 342-353 (2024). Chen, S. et al. Effects of articular disc or condylar cartilage resection on mandibular growth in young rats. Arch. Oral Biol. 97 , 67-71 (2019). Hayashi, H. et al. Role of articular disc in condylar regeneration of the mandible. Exp. Anim. 63 , 395-401 (2014). Lim, W. H., Choi, B., Lee, J. Y. & Ahn, S. J. Dentofacial characteristics in orthodontic patients with centric relation-maximum intercuspation discrepancy. Angle. Orthod. 84 , 939-945 (2014). Ahn, S. J., Lee, S. J. & Kim, T. W. Orthodontic effects on dentofacial morphology in women with bilateral TMJ disk displacement. Angle. Orthod. 77 , 288-295 (2007). Eduardo Charles Pagotto, L., Freitas de Morais, E., de Santana Santos, T. & Pires Pastore, G. Evaluating changes in the condylar head after orthognathic surgery with or without articular disc repositioning: a systematic review. Br. J. Oral. Maxillofac. Surg. 62 , 340-348 (2024). Sale, H., Bryndahl, F. & Isberg, A. Temporomandibular joints in asymptomatic and symptomatic nonpatient volunteers: a prospective 15-year follow-up clinical and MR imaging study. Radiol. 267 , 183-194 (2013). Meng, X., Liu, S., Wu, Z. & Guo, L. Application of panoramic radiographs in the diagnosis of temporomandibular disorders. Medicine. 103 , e36469 (2024). Durham, J. et al. Constructing the brief diagnostic criteria for temporomandibular disorders (bDC/TMD) for field testing. J. Oral. Rehabil. 51 , 785-794 (2024). Ahn, S.-J. et al. Relationships between temporomandibular joint disk displacements and condylar volume. Oral Surg. Oral Med. Oral Pathol. Oral Radiol. 125 , 192-198 (2018). Seo, B.-Y., An, J.-S., Chang, M.-S., Huh, K.-H. & Ahn, S.-J. Changes in condylar dimensions in temporomandibular joints with disk displacement. Oral Surg. Oral Med. Oral Pathol. Oral Radiol. 129 , 72-79 (2020). Seo, B.-Y., Huh, K.-H., An, J.-S., Chang, M.-S. & Ahn, S.-J. Relationship of computed tomography–verified degenerative condylar morphology with temporomandibular joint disk displacement and sex. Oral Surg. Oral Med. Oral Pathol. Oral Radiol. 132 , 93-103 (2021). Mani, S., Raghavan, S., Birur, P., Gurudath, S. & Keerthi, G. Accuracy of the detection of simulated condylar lesions on panoramic radiography and cone-beam computed tomography. J Adv. Clin. Res. Insights. 5 , 145-150 (2018). Ahn, S.-J., Kim, T.-W., Lee, D.-Y. & Nahm, D.-S. Evaluation of internal derangement of the temporomandibular joint by panoramic radiographs compared with magnetic resonance imaging. Am. J. Orthod. Dentofacial. Orthop. 129 , 479-485 (2006). Faghani, S. et al. Quantifying Uncertainty in Deep Learning of Radiologic Images. Radiol. 308 , e222217 (2023). Misera, L., Muller-Franzes, G., Truhn, D. & Kather, J. N. Weakly Supervised Deep Learning in Radiology. Radiol. 312 , e232085 (2024). Bai, G. et al. Automatic temporomandibular disc displacement diagnosis via deep learning. Displays. 77 , 102394 (2023). Jocher, G. et al. ultralytics/yolov5: v3. 0. Zenodo (2020). He, K., Zhang, X., Ren, S. & Sun, J. in Proceedings of the IEEE conference on computer vision and pattern recognition. 770-778 (2016) Fu, J. et al. in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 3146-3154 (2019) Pan, H., Han, H., Shan, S. & Chen, X. in Proceedings of the IEEE conference on computer vision and pattern recognition. 5285-5294 (2018) Goto, T. K. et al. Correlation of mandibular deviation with temporomandibular joint MR dimensions, MR disk position, and clinical symptoms. Oral Surg. Oral Med. Oral Pathol. Oral Radiol. 100 , 743-749 (2005). Tanaka, E., Detamore, M. & Mercuri, L. Degenerative disorders of the temporomandibular joint: etiology, diagnosis, and treatment. J. Dent. Res. 87 , 296-307 (2008). Dias, I. M., Coelho, P. R., Assis, N. M. S. P., Leite, F. P. P. & Devito, K. L. Evaluation of the correlation between disc displacements and degenerative bone changes of the temporomandibular joint by means of magnetic resonance images. Int. J. Oral. Maxillofac. Surg. 41 , 1051-1057 (2012). Silva, M. A. G. et al. Prevalence of degenerative disease in temporomandibular disorder patients with disc displacement: A systematic review and meta-analysis. J. Craniomaxillofac. Surg. 48 , 942-955 (2020). de Pontes, M. L. C., Melo, S. L. S., Bento, P. M., Campos, P. S. F. & de Melo, D. P. Correlation between temporomandibular joint morphometric measurements and gender, disk position, and condylar position. Oral Surg. Oral Med. Oral Pathol. Oral Radiol. 128 , 538-542 (2019). Kambylafkas, P., Murdock, E., Gilda, E., Tallents, R. H. & Kyrkanides, S. Validity of panoramic radiographs for measuring mandibular asymmetry. Angle. Orthod. 76 , 388-393 (2006). Arayapisit, T. et al. Understanding the mandibular condyle morphology on panoramic images: A cone beam computed tomography comparison study. Cranio. 41 , 354-361 (2023). Tables Table 1. Demographic information of study participants Dataset for ROI detection Dataset for TMJ DD classification Number of participants (Age range, years) Number of Participants (Age range, years) Number of condyles (%) Disk displacement status Total NR DDR DDNR Female 871 (4.0–63.2) 570 (15.0–47.3) 347 (22.2) 312 (19.9) 456 (29.2) 1115 (71.3) Male 598 (3.2–57.1) 229 (15.1–42.3) 218 (13.9) 110 (7.0) 121 (7.7) 449 (28.7) Total 1469 (3.2–63.2) 799 (15.0–47.3) 565 (36.1) 422 (27.0) 577 (36.9) 1564 (100) ROI, region of interest; NR, normal disk position; DDR, disk displacement with reduction; DDNR, disk displacement without reduction; TMJ, temporomandibular joint; DD, disk displacement; DD, disk displacement; NR, normal disk position, DDR, disk displacement with reduction; DDNR, disk displacement without reduction. Table 2. Multi-class classification performance for normal disk position (NR), disk displacement with reduction (DDR), and disk displacement without reduction (DDNR) Precision Recall F1-score Disk status NR 0.7143 0.8730 0.7857 DDR 0.5162 0.3810 0.4384 DDNR 0.8305 0.7903 0.8099 Overall performance Macro Avg 0.6870 0.6814 0.6780 Weighted Avg 0.7076 0.7186 0.7073 Accuracy 0.7186 Table 3. Confusion matrix and diagnostic performance for binary classification Confusion matrix Diagnostic performance Predicted Sensitivity Specificity Accuracy NR DD Actual NR 55 8 0.7885 0.8730 0.8204 DD 22 82 NR, normal disk position; DD, disk displacement. Disk displacement with reduction and disk displacement without reduction was treated as a single class for binary classification. Additional Declarations No competing interests reported. Supplementary Files Supplementaryfilefinal.docx Cite Share Download PDF Status: Published Journal Publication published 07 Jan, 2026 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 11 Aug, 2025 Reviews received at journal 03 Aug, 2025 Reviews received at journal 02 Aug, 2025 Reviews received at journal 28 Jul, 2025 Reviewers agreed at journal 27 Jul, 2025 Reviewers agreed at journal 26 Jul, 2025 Reviewers agreed at journal 25 Jul, 2025 Reviewers agreed at journal 25 Jul, 2025 Reviewers agreed at journal 25 Jul, 2025 Reviewers invited by journal 25 Jul, 2025 Editor invited by journal 10 Jul, 2025 Editor assigned by journal 08 Jul, 2025 Submission checks completed at journal 07 Jul, 2025 First submitted to journal 05 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-7053119","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":492176927,"identity":"3bac9319-ff10-4b79-9291-4374acbb85bb","order_by":0,"name":"Jong-Soo Kim","email":"","orcid":"","institution":"Seoul National University","correspondingAuthor":false,"prefix":"","firstName":"Jong-Soo","middleName":"","lastName":"Kim","suffix":""},{"id":492176929,"identity":"71df9f33-21c0-4410-9fa1-ddd957e483ee","order_by":1,"name":"Taehyeong Kim","email":"","orcid":"","institution":"Seoul National University","correspondingAuthor":false,"prefix":"","firstName":"Taehyeong","middleName":"","lastName":"Kim","suffix":""},{"id":492176930,"identity":"b1129ed1-e6dc-42b0-bc88-bfc68b726358","order_by":2,"name":"Yong-Hee Ahn","email":"","orcid":"","institution":"Seoul National University","correspondingAuthor":false,"prefix":"","firstName":"Yong-Hee","middleName":"","lastName":"Ahn","suffix":""},{"id":492176932,"identity":"c8cd22ba-dbb4-4fcc-ab09-e6ab0cf9ba2f","order_by":3,"name":"Hong-Gee Kim","email":"","orcid":"","institution":"Seoul National University","correspondingAuthor":false,"prefix":"","firstName":"Hong-Gee","middleName":"","lastName":"Kim","suffix":""},{"id":492176934,"identity":"d46d6a63-be97-4efd-bdcf-eb5ffeeb1dbe","order_by":4,"name":"Sug-Joon Ahn","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAApklEQVRIiWNgGAWjYDACZhBRAeMdIEIHDzNIzxmStICsYWwjRYs9O//BBx/n1eUZHGB++IHhzD2iHMZsOHPb4WKDA2zGEgw3ionSwibNu+1A4oYDDGYMDB8SiNUypw6ohf0bKVoamIFaeIC23CBGy2FmY8MZxw4nzjzMUyyRcIYILez9Bx8++FBTl9h3vH3jhw/HiNCCAKAoJUnDKBgFo2AUjALcAAAd+DOJqMgtOwAAAABJRU5ErkJggg==","orcid":"","institution":"Seoul National University","correspondingAuthor":true,"prefix":"","firstName":"Sug-Joon","middleName":"","lastName":"Ahn","suffix":""}],"badges":[],"createdAt":"2025-07-05 12:38:25","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7053119/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7053119/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-025-34657-1","type":"published","date":"2026-01-07T15:58:16+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":87898888,"identity":"1cb9cc03-4662-4ffc-a7d9-9216d52a0279","added_by":"auto","created_at":"2025-07-30 07:57:55","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":10097585,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart illustrating the screening, inclusion, and exclusion process for subjects with temporomandibular joint (TMJ) disk displacement: ROI, region of interest; DD, disk displacement.\u003c/p\u003e","description":"","filename":"Fig1.png","url":"https://assets-eu.researchsquare.com/files/rs-7053119/v1/5097492ddf7c0c44ef6480e5.png"},{"id":87897553,"identity":"cf18b3e2-a102-4fa8-97c2-bca7f6abab37","added_by":"auto","created_at":"2025-07-30 07:49:55","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":14206130,"visible":true,"origin":"","legend":"\u003cp\u003eAnnotation tool for marking anatomical landmarks on the condylar regions in orthopantomogram (a and b) and the orthopantomogram images representing different disk conditions (c-h). (a) Red dots indicate key anatomical landmarks on both left and right mandibular condyles. (b) Illustration of five anatomical landmarks used for annotation of region of interest, numbered in order: 1) the most prominent point of the condyle, 2) the most superior point of the condyle, 3) the superior point of the tangent to the posterior edge of the ramus, 4) the most inferior point of the tangent to the posterior border of the ramus, and 5) the most inferior point of the sigmoid notch. Normal disk position in the male (c) and female (d) condyles. Disk displacement with reduction in the male (e) and female (f) condyles. Disk displacement without reduction in the male (g) and female (h) condyles.\u003c/p\u003e","description":"","filename":"Fig2.png","url":"https://assets-eu.researchsquare.com/files/rs-7053119/v1/2d95bc76a920d59e8fce6f16.png"},{"id":87897549,"identity":"e5fe744e-da86-42a3-a20b-51f531ee38a7","added_by":"auto","created_at":"2025-07-30 07:49:55","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":6985502,"visible":true,"origin":"","legend":"\u003cp\u003eOverall architecture (a) and evaluation methodology (b) of TMJ-PanoNet for identifying disk displacement of the temporomandibular joint from the orthopantomogram. (a) The model consisted of two sequential stages: 1) detection of the region of interest using YOLOv5s to identify the condylar regions, and 2) classification of disk displacement using a combination of ResNet-18 and DANet. (b) Comprehensive evaluation methodology for TMJ-PanoNet. The flowchart outlines a three-pronged evaluation approach: 1) multi-class classification performance assessed via precision, recall, F1-score, and accuracy; 2) binary classification evaluation based on sensitivity, specificity, and accuracy; and 3) continuous estimation capability measured using the coefficient of determination (R²), Pearson correlation coefficient (r), mean squared error (MSE), and mean absolute error (MAE). NR, normal disk position; DDR, disk displacement with reduction; DDNR, disk displacement without reduction; DD, disk displacement; TMJ, temporomandibular joint.\u003c/p\u003e","description":"","filename":"Fig3.png","url":"https://assets-eu.researchsquare.com/files/rs-7053119/v1/5da21b99239582c6ad9a0aea.png"},{"id":87897548,"identity":"21616ab7-884d-4f7a-a0bb-de9c36a95427","added_by":"auto","created_at":"2025-07-30 07:49:55","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":3602577,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrix with counts and probabilities for classification of disk displacement of the temporomandibular joint (a) and correlation between actual and predicted values (b). (a) The confusion matrix shows the performance of the TMJ-PanoNet model in classifying normal disk position (NR), disk displacement with reduction (DDR), and disk displacement without reduction (DDNR) cases. Each cell contains the count of predictions and its corresponding percentage of the total samples. The diagonal elements represent correct classifications, while off-diagonal elements indicate misclassifications. (b) Correlation between magnetic resonance imaging (MRI)-verified disk displacement (DD) scores and predicted values. This scatter plot illustrates the relationship between the actual and predicted values by the TMJ-PanoNet model. Each point represents a single case, with the x-axis showing predicted values and the y-axis showing actual values. The red line indicates the best-fit linear regression. The model's performance was quantified by several metrics: the coefficient of determination (R²) of 0.4871 reflects how well the model's predictions match the actual values; the Pearson correlation coefficient (r) of 0.7531 suggests a strong positive correlation between actual and predicted values; and the mean squared error (MSE) of 0.3839 and mean absolute error (MAE) of 0.3272 provide measures of prediction accuracy. TMJ, temporomandibular joint.\u003c/p\u003e","description":"","filename":"Fig4.png","url":"https://assets-eu.researchsquare.com/files/rs-7053119/v1/ce20dcb6d57805e6b7ee74c7.png"},{"id":100070938,"identity":"918832f4-49f5-4f6d-8365-8cae01d70ebf","added_by":"auto","created_at":"2026-01-12 16:18:48","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":31806705,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7053119/v1/043a363c-4eca-4645-a62e-43e1667943e7.pdf"},{"id":87897546,"identity":"53c0649b-6d94-4d7b-8832-3fbb824f94ca","added_by":"auto","created_at":"2025-07-30 07:49:54","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":25479,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementaryfilefinal.docx","url":"https://assets-eu.researchsquare.com/files/rs-7053119/v1/20997c08ca62cb3158918649.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Two-stage deep learning approach for screening disk displacement of the temporomandibular joint using orthopantomogram","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe temporomandibular joint (TMJ) is configured to withstand repetitive and high functional loads between the temporal bone and the mandibular condyle. The articular disk, positioned between the condyle and temporal bone, serves as a stress distributor during jaw activity\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Consequently, disk displacement (DD) can increase stress on surrounding tissues, leading to damage to the articulating surfaces of the condyle and temporal bone\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Therefore, DD is associated with degenerative changes in TMJ articulating structures, manifesting as various symptoms such as joint noise, pain, and restricted mouth opening\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. In addition, TMJ DD influences aspects of the stomatognathic system, such as condylar and mandibular growth\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e, recovery and regeneration of the condyle\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e, and centric relation-maximum intercuspation discrepancy\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e, which may have a significant effect on the results of oral rehabilitation\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e,\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. Therefore, it is important to screen for TMJ DD prior to commencing oral rehabilitation.\u003c/p\u003e\u003cp\u003eMagnetic resonance imaging (MRI) is the gold standard for determining the position of the articular disk due to its superior ability to depict the disk\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e. However, its clinical use in dentistry is limited by cost-effectiveness concerns, spatial constraints, and various contraindications\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. While diagnostic criteria for temporomandibular disorders based on clinical variables provide an alternative, they require time-consuming training and examinations\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e. As a result, there is a strong need for simpler and more cost-effective methods to screen for TMJ DD.\u003c/p\u003e\u003cp\u003eA strong correlation between DD and degenerative condylar changes has been reported using MRI and computed tomography\u003csup\u003e\u003cspan additionalcitationids=\"CR13\" citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e–\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. Condyles affected by DD exhibit a higher prevalence of degenerative changes, such as erosion, osteophyte formation, and flattening, compared to those with normal disk positioning (NR)\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. These changes are associated with altered condylar size and shape, and joints with DD have reduced condylar width, height, length, and volume compared to those with NR\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. Orthopantomograms (OTGs) are widely used for the initial screening of condylar pathologies due to their low cost, minimal radiation exposure, and ability to provide a comprehensive view of the joint structure\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. Given that TMJ DD is associated with degenerative condylar changes and altered condylar morphologies\u003csup\u003e\u003cspan additionalcitationids=\"CR13\" citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e–\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e, OTG may serve as a viable tool for detecting DD-related structural changes of the mandibular condyle. A previous study also reported that specific condylar morphologies, such as decreased condylar height and a distally inclined condyle, were frequently observed in OTGs of patients with TMJ DD\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eRecently, deep learning approaches have been increasingly utilized in dentistry to improve the diagnosis of complex oral diseases\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e,\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. While TMJ DD screening models have been developed using MRI or clinical data\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e, the potential of deep learning applied to OTG images remains unexplored. The purpose of this study was 1) to develop a deep learning-based approach for screening TMJ DD using OTG and 2) to evaluate the performance of this approach.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cb\u003eStudy design\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWe analyzed data from patients who visited Seoul National University Dental Hospital between 1994 and 2023. The inclusion criteria were as follows: 1) undergoing OTG for a general examination, 2) no history of TMJ trauma, TMJ treatment, orthodontic treatment, or orthognathic surgery, and 3) no systemic disease, including rheumatoid arthritis and congenital anomalies. This retrospective study was approved by the Institutional Review Board of the University (S-D20220026). The requirement for informed consent was waived due to the retrospective nature of this study using existing records. This study was conducted in accordance with the Declaration of Helsinki.\u003c/p\u003e\u003cp\u003eWe proposed a two-stage deep learning approach, TMJ-PanoNet, for identifying TMJ DD using OTG. This approach utilized two distinct datasets: one for training a model to detect regions of interest (ROI) and another for TMJ DD identification. The ROI detection dataset included 1,469 OTG images from 598 males and 871 females who underwent only OTG. The TMJ DD identification dataset comprised 1,564 joints from 229 males and 570 females who underwent both OTG and TMJ MRI within 3 months. For the DD identification dataset, participants were limited to those aged 15–50 years, as active condylar growth during puberty can influence condylar size and shape\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e, and participants over 50 years were excluded due to a small sample size. The inclusion and exclusion process for screening for TMJ DD is shown in a flowchart in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003cb\u003eOTG image acquisition\u003c/b\u003e\u003c/p\u003e\u003cp\u003eOTG images were acquired using the OP-100 (Instrumentarium, Tuusula, Finland). Patients’ heads were positioned optimally using standard head positioning devices and bite blocks, following the manufacturer’s instructions. Imaging parameters were set to 70 kVp, 10 mA, and an exposure time of 17.6 s. OTG data were stored in TIF format, either by scanning analogue images or downloading digital images from a web-based picture archiving and communication system.\u003c/p\u003e\u003cp\u003e\u003cb\u003eDisk position determination\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTMJ MRI was performed using Signa Horizon (GE, Waukesha, WI) or Magnetom Vita (Siemens, Erlangen, Germany) (\u003cb\u003eSupplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e online\u003c/b\u003e). Reference annotations for DD categories were provided by two radiologists with over 20 years of experience in TMJ MRI interpretation, who independently evaluated disk positions in a blinded manner. Following a thorough assessment, DD was classified into three categories: normal disk position (NR), DD with reduction (DDR), and DD without reduction (DDNR) as previously described\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e,\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. Any discrepancies in disk position classification were resolved by consensus. Condyles with partial DD, DD with partial reduction, pure-sideway DD, and posterior DD were excluded due to their small sample sizes.\u003c/p\u003e\u003cp\u003eFor continuous severity assessment, DD severity scores were assigned based on clinical progression: NR was assigned a score of 0 (no displacement), DDR a score of 1 (intermediate severity), and DDNR a score of 2 (most severe condition).\u003c/p\u003e\u003cp\u003e\u003cb\u003eDetermination of ROI in OTG Images\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWe developed a specialized annotation tool for precise identification of anatomical landmarks to facilitate accurate detection of ROI around the mandibular condyle in OTG images (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea). The tool’s interface included options to navigate through images, clear points, and save annotations, facilitating efficient and consistent landmark identification across multiple images. Five anatomical landmarks were identified in each region (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb) to define the ROI around the condylar area associated with DD (Figs.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec-h). An experienced orthodontist with over 25 years of practice carefully marked these landmarks, which directly established the ROI boundaries across all images. The annotated regions then served as ground truth data for training the ROI detection model in the first stage of TMJ-PanoNet.\u003c/p\u003e\u003cp\u003e\u003cb\u003eDataset partitioning for model development\u003c/b\u003e\u003c/p\u003e\u003cp\u003ePrevious studies have shown that the relationship between condylar morphology and DD is not significantly influenced by sex\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e,\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e; therefore, data from both males and females were combined.\u003c/p\u003e\u003cp\u003eFor model training and evaluation, both datasets were divided into training, validation, and testing subsets. The ROI detection dataset of 1,469 OTG images was randomly split into training (1,028 images), validation (293 images), and testing (148 images) subsets. The DD classification dataset was also divided into training (1,244 joints), validation (153 joints), and testing (167 joints) subsets. These two datasets were completely separate with no patient overlapping. Additionally, for the DD classification dataset, patient-level separation was implemented to prevent data leakage, ensuring that both condyles from the same patient were assigned to the same subset. This rigorous segregation ensured that the testing dataset consisted of completely unseen patients, reflecting model performance on new cases and mimicking real clinical scenarios.\u003c/p\u003e\u003cp\u003e\u003cb\u003eArchitecture of TMJ-PanoNet\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTMJ-PanoNet was designed as a two-stage deep learning model (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea). The first stage focused on the ROI detection, whereas the second stage analyzed these ROIs to classify DD status. In the first stage, YOLOv5s\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e was trained to automatically detect and localize the ROIs for both the left and right sides of the TMJ using the annotated regions in OTG images. This approach enabled effective localization across patients with varying mandibular structures, regardless of the variations in scales and orientations of the mandibular condyles. Before passing the ROIs to the second stage for TMJ DD classification, the left condylar region was horizontally flipped to match the orientation of the right condylar region. Additionally, all detected ROIs were resized to a fixed resolution (\u003cem\u003ee.g.\u003c/em\u003e, 224×224 pixels). This standardization ensured uniform orientation across samples and normalized condylar features for consistent analysis. Furthermore, this approach improved data homogeneity by treating left and right condyles identically, which helped address the challenge of limited data availability and enhanced the model’s generalization capability.\u003c/p\u003e\u003cp\u003eThe second stage, operating on these detected ROIs, performed classification of TMJ DD using a pretrained ResNet-18\u003csup\u003e21\u003c/sup\u003e for feature extraction, which was fine-tuned on our dataset, combined with position and channel attention modules of DANet\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e to emphasize crucial features. ResNet-18 was chosen due to its balance between efficiency and performance, with fewer parameters compared to deeper variants to mitigate overfitting risk, while maintaining effective feature learning through its residual connections. The DANet’s position and channel attention modules enhanced the identification of crucial anatomical features in the condylar regions (Figs.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec-h). For model training, we utilized a hybrid loss function approach by combining mean-variance loss\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e with standard cross-entropy loss. The feature maps were then processed through fully-connected layers (256 and 128 neurons) with batch normalization and dropout (p = 0.5) for regularization. Detailed model training procedures and implementation details are provided in the Supplementary Materials.\u003c/p\u003e\u003cp\u003e\u003cb\u003eTwo-stage inference process\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTMJ-PanoNet performed inference in two sequential stages. First, the model automatically extracted and standardized ROIs from OTG, creating uniform condylar regions despite variations in TMJ scale. The standardized orientation of condylar regions enabled independent assessment of left and right joints while maintaining consistent feature analysis. The classification model then processed these standardized inputs to estimate TMJ DD severity scores. The severity score for each sample \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:i\\)\u003c/span\u003e\u003c/span\u003e was computed as the weighted average of class labels using predicted probabilities:\u003c/p\u003e\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"384\" height=\"84\"\u003e\u003c/p\u003e\u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:K=3\\)\u003c/span\u003e\u003c/span\u003e is the total number of classes, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:j\\)\u003c/span\u003e\u003c/span\u003e represents the class labels (NR = 0, DDR = 1, DDNR = 2), and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{p}_{i,j}\\)\u003c/span\u003e\u003c/span\u003e denotes the predicted probability that sample \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:i\\)\u003c/span\u003e\u003c/span\u003e belongs to class \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:j\\)\u003c/span\u003e\u003c/span\u003e. These continuous scores (ranging from 0 to 2) were converted to categorical classifications using thresholds: NR (0-0.5), DDR (0.5–1.5), and DDNR (1.5-2). This method enabled both continuous severity assessment and categorical classification for comprehensive evaluation of TMJ DD status.\u003c/p\u003e\u003cp\u003e\u003cb\u003eModel evaluation\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe performance of TMJ-PanoNet was evaluated using an independent test set of 167 joints, representing approximately 10.7% of the total 1,564 joints in the second dataset. We conducted multi-class (NR, DDR, and DDNR) and binary classification (NR and DD) analyses, alongside DD severity classification, to assess the model's classification capabilities (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb). Model performance metrics included precision, recall, F1 score, and accuracy for multi-class evaluation, and sensitivity, specificity, and accuracy for binary classification. For continuous severity assessment, the coefficient of determination (R²), Pearson correlation coefficient (r), Mean Squared Error (MSE), and Mean Absolute Error (MAE) were calculated (detailed information is available in the Supplementary Material). All metrics were computed using Python version 3.8.10 with scikit-learn library version 1.2.0. Statistical analysis was performed by an expert with more than 15 years of experience in machine learning and image processing.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003ePatient characteristics of this experiment\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe inclusion and exclusion process for TMJ DD screening is shown in a flowchart in \u003cstrong\u003eFig. 1\u003c/strong\u003e. After applying the exclusion criteria, the final ROI detection dataset included 1,469 images (871 females and 598 males; average age 14.52 ± 7.78 years). For the TMJ DD classification dataset, 1,564 joints from 799 patients (570 females and 229 males; average age 23.98 ± 5.94 years) were included. Among these, 565 joints (36.1%) had NR, 422 joints (27.0%) had DDR, and 577 joints (36.9%) had DDNR (\u003cstrong\u003eTable 1\u003c/strong\u003e).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWhen we attempted to train and evaluate separate models based on sex and age groups, these approaches did not result in significant improvements (data not shown). This may be because condylar changes associated with DD are not significantly influenced by sex-related factors after active growth\u003csup\u003e12-14\u003c/sup\u003e. As a result, male and female condyles aged 15–50 years were combined to increase data volume and diversity, thereby enhancing the model.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDiagnostic performance of ROI detection\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe ROI detection module achieved perfect detection performance with a 100% success rate (intersection over union (IoU) ≥ 0.5 across all 293 test images), localizing condylar regions in every case. This high performance can be attributed to the relatively straightforward nature of the ROI detection task, as mandibular condyles have consistent anatomical locations within OTG images. Based on these accurately detected ROIs, the diagnostic performance for DD classification was evaluated.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDiagnostic performance of multi-class classification\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFig. 4a\u003c/strong\u003e presents the confusion matrix for the three-class classification (NR, DDR, and DDNR). In the multi-class analysis (\u003cstrong\u003eTable 2\u003c/strong\u003e), our model achieved an overall accuracy of 71.86% (120/167), with varying performance across categories. Particularly, the model demonstrated strong performance in identifying NR (precision: 71.43%, recall: 87.30%, and F1-score: 78.57%) and DDNR cases (precision: 83.05%, recall: 79.03%, and F1-score: 80.99%), indicating reliable detection of both normal and the most severe conditions. However, it showed relatively lower performance for DDR cases (precision: 51.62%, recall: 38.10%, F1-score: 43.84%), suggesting challenges in identifying this intermediate condition. The model's strength in discriminating between extreme cases was particularly evident, with misclassifications between NR and DDNR occurring in only six cases (3.59%) of the total elements in the confusion matrix elements.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDiagnostic performance of binary classification\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo evaluate the model's performance in a clinical screening context, we conducted a binary classification analysis, where DDR and DDNR were combined into a single DD class (\u003cstrong\u003eTable 3\u003c/strong\u003e). This simplified classification showed improved performance with an accuracy of 82.04% (137/167). The model correctly identified 55 of 63 NR cases and 82 of 104 DD cases, achieving a sensitivity of 78.85% (82/104) and a specificity of 87.30% (55/63). These results suggest that the model could be particularly useful as an initial screening for detecting the presence of DD.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDiagnostic performance of continuous severity estimation\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eBeyond categorical classification, we assessed the model's ability to estimate continuous TMJ DD severity (\u003cstrong\u003eFig. 4b\u003c/strong\u003e), which could provide more nuanced information about disease progression. The analysis yielded an R² of 0.4871, demonstrating the model's moderate predictive power for continuous severity estimation. The Pearson r of 0.7531 demonstrated a strong positive correlation between predicted and actual values, suggesting consistent alignment between the model's estimates and clinical assessments. The model achieved an MSE of 0.3839 and an MAE of 0.3272, with these error metrics indicating reasonable prediction accuracy considering the subjective nature of severity assessment. This continuous severity estimation capability could be particularly valuable for monitoring disease progression over time.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eSignificant efforts have been made to develop a simple and convenient tool for screening TMJ DD to aid in proper treatment planning for patients with potential TMJ disorders\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e,\u003cspan additionalcitationids=\"CR25 CR26\" citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e. While MRI remains the gold standard for diagnosing TMJ DD\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e, its clinical use in dentistry is limited due to various constraints\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. Previous studies have demonstrated that DD is frequently accompanied by degenerative changes in the mandibular condyle\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e, resulting in alterations in its size and shape\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e,\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. Given that OTG is a simple and cost-effective imaging tool capable of providing a comprehensive view of the condyle and its surrounding structures, it may hold potential for detecting joint morphology changes associated with DD. To the best of our knowledge, this is the first study to identify TMJ DD using OTG combined with deep learning.\u003c/p\u003e\u003cp\u003eTMJ-PanoNet classified TMJ DD into three categories\u0026mdash;NR, DDR, and DDNR\u0026mdash;with an overall accuracy of 71.86% (120/167) (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). However, its ability to identify DDR was lower than its performance in identifying NR and DDNR (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). This discrepancy may be attributed to differences in condylar morphology among the three DD categories. In DDR, the displaced disk returns to its normal position during mouth opening, restoring disk function. As a result, condylar changes associated with DD are less pronounced in DDR than in DDNR\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e,\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e. In this study, differences in condylar morphology were less distinct between NR and DDR than between NR and DDNR or DDR and DDNR, regardless of sex (Figs.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC-H). Since OTG has inherent limitations in detecting minor morphological changes, subtle differences between NR (Figs.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC-D) and DDR (Figs.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eE-F) may be challenging to identify in the TMJ-PanoNet.\u003c/p\u003e\u003cp\u003eNonetheless, some morphological changes were observed in DDR cases, despite their lower severity. Specifically, the condylar width is initially affected by the resorption of the lateral pole before more extensive morphological changes occur as DD progresses\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e,\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. Since OTG primarily captures the lateral and medial thirds of the condylar head\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e, early changes in condylar width at the lateral pole may aid in DDR identification, which explains TMJ-PanoNet's ability to detect DDR despite its lower F1 score (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eIn contrast to DDR, DDNR cases do not exhibit disk reduction upon mouth opening, resulting in total loss of disk function during jaw function. This loss is accompanied by more pronounced degenerative changes in the condylar surfaces. Consequently, radiographically detectable degenerative changes are more evident in DDNR than in DDR\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e, with DDNR cases exhibiting the smallest condyles and the most severe morphological alterations\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. In this study, significant differences in size and shape were observed between NR (Figs.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC-D) and DDNR (Figs.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eG-H), which may explain the minimal misclassification between these extreme categories (only six cases, 3.59% of the total). Given that DDNR cases have a higher likelihood of developing osteoarthritis compared to NR and DDR cases\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e, TMJ-PanoNet\u0026rsquo;s superior ability to detect DDNR may assist clinicians in diagnosing and planning treatments for patients with TMJ DD.\u003c/p\u003e\u003cp\u003eWhen DDR and DDNR were treated as a single category, the model achieved improved classification performance (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). Specifically, the accuracy of binary classification (82.04%) was approximately 10% higher than that of the three-class classification (71.86%). The model\u0026rsquo;s accuracy, sensitivity, and specificity are around 80%, suggesting that TMJ-PanoNet may be useful for the initial screening of TMJ DD in clinical settings.\u003c/p\u003e\u003cp\u003eBeyond condylar changes, the morphology of the articular eminence and glenoid fossa is also significantly influenced by TMJ DD. A flattened articular eminence and a larger glenoid fossa have been associated with TMJ DD\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e. In this study, the posterior slope of the articular eminence and the glenoid fossa were included in the ROI (Figs.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC-H), potentially contributing to the model\u0026rsquo;s improved performance in identifying DD using OTG.\u003c/p\u003e\u003cp\u003eAdditionally, TMJ-PanoNet demonstrated potential for estimating TMJ DD severity on a continuous scale (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb). The model achieved an R\u0026sup2; of 0.4871, a Pearson correlation coefficient of 0.7531, an MSE of 0.3839, and an MAE of 0.3272. These results indicate a moderate to strong correlation between predicted and actual values, highlighting the model\u0026rsquo;s ability to capture subtle differences in DD severity with reasonable accuracy.\u003c/p\u003e\u003cp\u003eThis study has some limitations. First, our dataset was derived from a single OTG device, restricting data variability. To ensure the model's generalizability, future studies should incorporate data from multiple OTG devices. However, previous research has shown that standardized head positioning and bite blocks yield acceptable reproducibility of angular and linear dimensions, regardless of the OTG device used\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e. Moreover, condylar morphology associated with TMJ DD is not dependent on OTG machines\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e,\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e, suggesting that TMJ-PanoNet\u0026rsquo;s predictive performance may not be significantly influenced by device variations. Second, our study population was limited to individuals aged 15 to 50 years. While this age group represents a significant portion of patients with TMJ DD in dental clinics, further studies including a broader age range and data from multiple OTG devices are necessary to enhance the model\u0026rsquo;s clinical applicability.\u003c/p\u003e\u003cp\u003eTMJ-PanoNet achieved a diagnostic accuracy of 71.86% in the three-class classification (NR, DDR, and DDNR). However, its performance significantly improved, reaching 82.04% accuracy, when DDR and DDNR were treated as a single category. Given its strong ability to differentiate between NR and DD in a binary classification setting, TMJ-PanoNet has the potential to serve as an initial screening tool for TMJ DD in clinical practice. The model may be particularly beneficial for clinicians who are not familiar with TMJ MRI, providing a simple, fast, and convenient decision-support system.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the New Faculty Startup Fund from\u0026nbsp;Seoul\u0026nbsp;National\u0026nbsp;University and Seoul\u0026nbsp;National\u0026nbsp;University Dental Hospital (02-2025-0508).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eConceptualization: S.J.A., T.K., H.G.K.. Writing—original draft: J.S.K., T.K., S.J.A.. Writing—revision, review \u0026amp; editing: J.S.K., T.K., Y.H.A., H.G.K., S.J.A. Formal analysis: J.S.K., T.K. Funding acquisition: T.K., S.J.A. Investigation: J.S.K., T.K. Supervision: Y.H.A., H.G.K., S.J.A. Validation: T.K., Y.H.A., H.G.K., S.J.A. Project administration: T.K., H.G.K., S.J.A. All authors reviewed the manuscript.\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\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eKuroda, S.\u003cem\u003e et al.\u003c/em\u003e Biomechanical and biochemical characteristics of the mandibular condylar cartilage. \u003cem\u003eOsteoarthritis. Cartilage.\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 1408-1415 (2009).\u003c/li\u003e\n\u003cli\u003eLee, Y.-H., Won, J. H., Kim, S., Auh, Q.-S. \u0026amp; Noh, Y.-K. Advantages of deep learning with convolutional neural network in detecting disc displacement of the temporomandibular joint in magnetic resonance imaging. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 11352 (2022).\u003c/li\u003e\n\u003cli\u003eMaracci, L. M.\u003cem\u003e et al.\u003c/em\u003e Influence of examiner calibration on clinical and mri diagnosis of temporomandibular joint disc displacement: a systematic review and meta-analysis. \u003cem\u003eDentomaxillofac. Radiol.\u003c/em\u003e \u003cstrong\u003e53\u003c/strong\u003e, 342-353 (2024).\u003c/li\u003e\n\u003cli\u003eChen, S.\u003cem\u003e et al.\u003c/em\u003e Effects of articular disc or condylar cartilage resection on mandibular growth in young rats. \u003cem\u003eArch. Oral Biol.\u003c/em\u003e \u003cstrong\u003e97\u003c/strong\u003e, 67-71 (2019).\u003c/li\u003e\n\u003cli\u003eHayashi, H.\u003cem\u003e et al.\u003c/em\u003e Role of articular disc in condylar regeneration of the mandible. \u003cem\u003eExp. Anim.\u003c/em\u003e \u003cstrong\u003e63\u003c/strong\u003e, 395-401 (2014).\u003c/li\u003e\n\u003cli\u003eLim, W. H., Choi, B., Lee, J. Y. \u0026amp; Ahn, S. J. Dentofacial characteristics in orthodontic patients with centric relation-maximum intercuspation discrepancy. \u003cem\u003eAngle. Orthod.\u003c/em\u003e \u003cstrong\u003e84\u003c/strong\u003e, 939-945 (2014).\u003c/li\u003e\n\u003cli\u003eAhn, S. J., Lee, S. J. \u0026amp; Kim, T. W. Orthodontic effects on dentofacial morphology in women with bilateral TMJ disk displacement. \u003cem\u003eAngle. Orthod.\u003c/em\u003e \u003cstrong\u003e77\u003c/strong\u003e, 288-295 (2007).\u003c/li\u003e\n\u003cli\u003eEduardo Charles Pagotto, L., Freitas de Morais, E., de Santana Santos, T. \u0026amp; Pires Pastore, G. Evaluating changes in the condylar head after orthognathic surgery with or without articular disc repositioning: a systematic review. \u003cem\u003eBr. J. Oral. Maxillofac. Surg.\u003c/em\u003e \u003cstrong\u003e62\u003c/strong\u003e, 340-348 (2024).\u003c/li\u003e\n\u003cli\u003eSale, H., Bryndahl, F. \u0026amp; Isberg, A. Temporomandibular joints in asymptomatic and symptomatic nonpatient volunteers: a prospective 15-year follow-up clinical and MR imaging study. \u003cem\u003eRadiol.\u003c/em\u003e \u003cstrong\u003e267\u003c/strong\u003e, 183-194 (2013).\u003c/li\u003e\n\u003cli\u003eMeng, X., Liu, S., Wu, Z. \u0026amp; Guo, L. Application of panoramic radiographs in the diagnosis of temporomandibular disorders. \u003cem\u003eMedicine.\u003c/em\u003e \u003cstrong\u003e103\u003c/strong\u003e, e36469 (2024).\u003c/li\u003e\n\u003cli\u003eDurham, J.\u003cem\u003e et al.\u003c/em\u003e Constructing the brief diagnostic criteria for temporomandibular disorders (bDC/TMD) for field testing. \u003cem\u003eJ. Oral. Rehabil.\u003c/em\u003e \u003cstrong\u003e51\u003c/strong\u003e, 785-794 (2024).\u003c/li\u003e\n\u003cli\u003eAhn, S.-J.\u003cem\u003e et al.\u003c/em\u003e Relationships between temporomandibular joint disk displacements and condylar volume. \u003cem\u003eOral Surg. Oral Med. Oral Pathol. Oral Radiol.\u003c/em\u003e \u003cstrong\u003e125\u003c/strong\u003e, 192-198 (2018).\u003c/li\u003e\n\u003cli\u003eSeo, B.-Y., An, J.-S., Chang, M.-S., Huh, K.-H. \u0026amp; Ahn, S.-J. Changes in condylar dimensions in temporomandibular joints with disk displacement. \u003cem\u003eOral Surg. Oral Med. Oral Pathol. Oral Radiol.\u003c/em\u003e \u003cstrong\u003e129\u003c/strong\u003e, 72-79 (2020).\u003c/li\u003e\n\u003cli\u003eSeo, B.-Y., Huh, K.-H., An, J.-S., Chang, M.-S. \u0026amp; Ahn, S.-J. Relationship of computed tomography\u0026ndash;verified degenerative condylar morphology with temporomandibular joint disk displacement and sex. \u003cem\u003eOral Surg. Oral Med. Oral Pathol. Oral Radiol.\u003c/em\u003e \u003cstrong\u003e132\u003c/strong\u003e, 93-103 (2021).\u003c/li\u003e\n\u003cli\u003eMani, S., Raghavan, S., Birur, P., Gurudath, S. \u0026amp; Keerthi, G. Accuracy of the detection of simulated condylar lesions on panoramic radiography and cone-beam computed tomography. \u003cem\u003eJ Adv. Clin. Res. Insights.\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, 145-150 (2018).\u003c/li\u003e\n\u003cli\u003eAhn, S.-J., Kim, T.-W., Lee, D.-Y. \u0026amp; Nahm, D.-S. Evaluation of internal derangement of the temporomandibular joint by panoramic radiographs compared with magnetic resonance imaging. \u003cem\u003eAm. J. Orthod. Dentofacial. Orthop.\u003c/em\u003e \u003cstrong\u003e129\u003c/strong\u003e, 479-485 (2006).\u003c/li\u003e\n\u003cli\u003eFaghani, S.\u003cem\u003e et al.\u003c/em\u003e Quantifying Uncertainty in Deep Learning of Radiologic Images. \u003cem\u003eRadiol.\u003c/em\u003e \u003cstrong\u003e308\u003c/strong\u003e, e222217 (2023).\u003c/li\u003e\n\u003cli\u003eMisera, L., Muller-Franzes, G., Truhn, D. \u0026amp; Kather, J. N. Weakly Supervised Deep Learning in Radiology. \u003cem\u003eRadiol.\u003c/em\u003e \u003cstrong\u003e312\u003c/strong\u003e, e232085 (2024).\u003c/li\u003e\n\u003cli\u003eBai, G.\u003cem\u003e et al.\u003c/em\u003e Automatic temporomandibular disc displacement diagnosis via deep learning. \u003cem\u003eDisplays.\u003c/em\u003e \u003cstrong\u003e77\u003c/strong\u003e, 102394 (2023).\u003c/li\u003e\n\u003cli\u003eJocher, G.\u003cem\u003e et al.\u003c/em\u003e ultralytics/yolov5: v3. 0. \u003cem\u003eZenodo\u003c/em\u003e (2020).\u003c/li\u003e\n\u003cli\u003eHe, K., Zhang, X., Ren, S. \u0026amp; Sun, J. in \u003cem\u003eProceedings of the IEEE conference on computer vision and pattern recognition.\u003c/em\u003e 770-778 (2016)\u003c/li\u003e\n\u003cli\u003eFu, J.\u003cem\u003e et al.\u003c/em\u003e in \u003cem\u003eProceedings of the IEEE/CVF conference on computer vision and pattern recognition.\u003c/em\u003e 3146-3154 (2019)\u003c/li\u003e\n\u003cli\u003ePan, H., Han, H., Shan, S. \u0026amp; Chen, X. in \u003cem\u003eProceedings of the IEEE conference on computer vision and pattern recognition.\u003c/em\u003e 5285-5294 (2018)\u003c/li\u003e\n\u003cli\u003eGoto, T. K.\u003cem\u003e et al.\u003c/em\u003e Correlation of mandibular deviation with temporomandibular joint MR dimensions, MR disk position, and clinical symptoms. \u003cem\u003eOral Surg. Oral Med. Oral Pathol. Oral Radiol. \u003c/em\u003e\u003cstrong\u003e100\u003c/strong\u003e, 743-749 (2005).\u003c/li\u003e\n\u003cli\u003eTanaka, E., Detamore, M. \u0026amp; Mercuri, L. Degenerative disorders of the temporomandibular joint: etiology, diagnosis, and treatment. \u003cem\u003eJ. Dent. Res.\u003c/em\u003e \u003cstrong\u003e87\u003c/strong\u003e, 296-307 (2008).\u003c/li\u003e\n\u003cli\u003eDias, I. M., Coelho, P. R., Assis, N. M. S. P., Leite, F. P. P. \u0026amp; Devito, K. L. Evaluation of the correlation between disc displacements and degenerative bone changes of the temporomandibular joint by means of magnetic resonance images. \u003cem\u003eInt. J. Oral. Maxillofac. Surg.\u003c/em\u003e \u003cstrong\u003e41\u003c/strong\u003e, 1051-1057 (2012).\u003c/li\u003e\n\u003cli\u003eSilva, M. A. G.\u003cem\u003e et al.\u003c/em\u003e Prevalence of degenerative disease in temporomandibular disorder patients with disc displacement: A systematic review and meta-analysis. \u003cem\u003eJ. Craniomaxillofac. Surg.\u003c/em\u003e \u003cstrong\u003e48\u003c/strong\u003e, 942-955 (2020).\u003c/li\u003e\n\u003cli\u003ede Pontes, M. L. C., Melo, S. L. S., Bento, P. M., Campos, P. S. F. \u0026amp; de Melo, D. P. Correlation between temporomandibular joint morphometric measurements and gender, disk position, and condylar position. \u003cem\u003eOral Surg. Oral Med. Oral Pathol. Oral Radiol. \u003c/em\u003e\u003cstrong\u003e128\u003c/strong\u003e, 538-542 (2019).\u003c/li\u003e\n\u003cli\u003eKambylafkas, P., Murdock, E., Gilda, E., Tallents, R. H. \u0026amp; Kyrkanides, S. Validity of panoramic radiographs for measuring mandibular asymmetry. \u003cem\u003eAngle. Orthod.\u003c/em\u003e \u003cstrong\u003e76\u003c/strong\u003e, 388-393 (2006).\u003c/li\u003e\n\u003cli\u003eArayapisit, T.\u003cem\u003e et al.\u003c/em\u003e Understanding the mandibular condyle morphology on panoramic images: A cone beam computed tomography comparison study. \u003cem\u003eCranio.\u003c/em\u003e \u003cstrong\u003e41\u003c/strong\u003e, 354-361 (2023).\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003e\u003cstrong\u003eTable\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e1.\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eDemographic information of study participants\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 20px;\"\u003e\n \u003cp\u003eDataset for ROI detection\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"5\" style=\"width: 69px;\"\u003e\n \u003cp\u003eDataset for TMJ DD classification\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\" style=\"width: 20px;\"\u003e\n \u003cp\u003eNumber of participants\u003c/p\u003e\n \u003cp\u003e(Age range, years)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"3\" style=\"width: 20px;\"\u003e\n \u003cp\u003eNumber of Participants\u003c/p\u003e\n \u003cp\u003e(Age range, years)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"4\" style=\"width: 48px;\"\u003e\n \u003cp\u003eNumber of condyles (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" style=\"width: 34px;\"\u003e\n \u003cp\u003eDisk displacement status\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eTotal\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003eNR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003eDDR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003eDDNR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003eFemale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 20px;\"\u003e\n \u003cp\u003e871\u003c/p\u003e\n \u003cp\u003e(4.0\u0026ndash;63.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 20px;\"\u003e\n \u003cp\u003e570\u003c/p\u003e\n \u003cp\u003e(15.0\u0026ndash;47.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e347 (22.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e312 (19.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 12px;\"\u003e\n \u003cp\u003e456 (29.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e1115 (71.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003eMale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 20px;\"\u003e\n \u003cp\u003e598\u003c/p\u003e\n \u003cp\u003e(3.2\u0026ndash;57.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 20px;\"\u003e\n \u003cp\u003e229\u003c/p\u003e\n \u003cp\u003e(15.1\u0026ndash;42.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e218 (13.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e110 (7.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 12px;\"\u003e\n \u003cp\u003e121 (7.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e449 (28.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003eTotal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 20px;\"\u003e\n \u003cp\u003e1469\u003c/p\u003e\n \u003cp\u003e(3.2\u0026ndash;63.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 20px;\"\u003e\n \u003cp\u003e799\u003c/p\u003e\n \u003cp\u003e(15.0\u0026ndash;47.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e565 (36.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e422 (27.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 12px;\"\u003e\n \u003cp\u003e577 (36.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e1564 (100)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eROI, region of interest; NR, normal disk position; DDR, disk displacement with reduction; DDNR, disk displacement without reduction; TMJ, temporomandibular joint; DD, disk displacement; DD, disk displacement; NR, normal disk position, DDR, disk displacement with reduction; DDNR, disk displacement without reduction.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e2.\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eMulti-class classification performance for normal disk position (NR), disk displacement with reduction (DDR), and disk displacement without reduction (DDNR)\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"602\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 142px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003eF1-score\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\" style=\"width: 142px;\"\u003e\n \u003cp\u003eDisk status\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003eNR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e0.7143\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e0.8730\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e0.7857\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003eDDR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e0.5162\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e0.3810\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e0.4384\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003eDDNR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e0.8305\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e0.7903\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e0.8099\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\" style=\"width: 142px;\"\u003e\n \u003cp\u003eOverall performance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003eMacro Avg\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e0.6870\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e0.6814\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e0.6780\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003eWeighted Avg\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e0.7076\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e0.7186\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e0.7073\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 115px;\"\u003e\n \u003cp\u003e0.7186\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3.\u0026nbsp;\u003c/strong\u003eConfusion matrix and diagnostic performance for binary classification\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"603\" class=\"fr-table-selection-hover\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"4\" style=\"width: 293px;\"\u003e\n \u003cp\u003eConfusion matrix\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" style=\"width: 310px;\"\u003e\n \u003cp\u003eDiagnostic performance\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 161px;\"\u003e\n \u003cp\u003ePredicted\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 103px;\"\u003e\n \u003cp\u003eSensitivity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 103px;\"\u003e\n \u003cp\u003eSpecificity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 103px;\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 85px;\"\u003e\n \u003cp\u003eNR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003eDD\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 80px;\"\u003e\n \u003cp\u003eActual\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eNR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 85px;\"\u003e\n \u003cp\u003e55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.7885\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8730\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.8204\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 53px;\"\u003e\n \u003cp\u003eDD\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 85px;\"\u003e\n \u003cp\u003e22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e82\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eNR, normal disk position; DD, disk displacement. Disk displacement with reduction and disk displacement without reduction was treated as a single class for binary classification.\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":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Temporomandibular joint, Disk displacement, Screening, Orthopantomogram ","lastPublishedDoi":"10.21203/rs.3.rs-7053119/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7053119/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThis study aimed to develop a deep learning-based approach for the initial screening of disk displacement (DD) of the temporomandibular joint (TMJ) using orthopantomograms (OTG). A two-stage deep learning model was proposed: first, regions of interest were detected from 1,469 orthopantomogram images using YOLOv5s; second, DD status was identified from 1,564 joints with the magnetic resonance imaging-verified DD status using ResNet-18 and DANet. Diagnostic performance was evaluated through multi-class and binary classification analyses. In three-class classification (normal disk position, DD with reduction, and DD without reduction), the model achieved an overall accuracy of 71.86%. It performed well in identifying normal disk position (F1 score: 78.57%) and DD without reduction (F1 score: 80.99%) but showed lower performance in detecting DD with reduction (F1 score: 43.84%). In the binary classification, where DD with reduction and DD without reduction were combined into a single class, the model demonstrated improved accuracy (82.04%), sensitivity (78.85%), and specificity (87.30%). The model exhibited potential in estimating DD severity (R\u0026sup2; = 0.4871). Given its strong ability to differentiate between normal disk position and DD in a binary classification setting, this tool has the potential to serve as an initial screening tool for TMJ DD in clinical practice.\u003c/p\u003e","manuscriptTitle":"Two-stage deep learning approach for screening disk displacement of the temporomandibular joint using orthopantomogram","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-30 07:49:50","doi":"10.21203/rs.3.rs-7053119/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-08-11T08:53:08+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-03T11:05:18+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-02T21:16:10+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-07-28T18:45:21+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"73497062706049587253891143926149633678","date":"2025-07-27T06:35:35+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"152450364509297668225409547576403962651","date":"2025-07-26T10:27:16+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"68702283220824262291589189953935904539","date":"2025-07-25T10:30:30+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"244034685796523140596337052223772984971","date":"2025-07-25T08:29:49+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"146045558065733544240406087835232721973","date":"2025-07-25T07:05:56+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-07-25T06:33:08+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-07-10T13:40:23+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-07-09T03:14:25+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-07-07T07:02:31+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-07-05T12:35:39+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"9afa75ec-070d-4987-be9e-ac77928e90a2","owner":[],"postedDate":"July 30th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":52256973,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":52256974,"name":"Health sciences/Diseases"},{"id":52256975,"name":"Health sciences/Health care"},{"id":52256976,"name":"Health sciences/Medical research"}],"tags":[],"updatedAt":"2026-01-12T16:15:35+00:00","versionOfRecord":{"articleIdentity":"rs-7053119","link":"https://doi.org/10.1038/s41598-025-34657-1","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2026-01-07 15:58:16","publishedOnDateReadable":"January 7th, 2026"},"versionCreatedAt":"2025-07-30 07:49:50","video":"","vorDoi":"10.1038/s41598-025-34657-1","vorDoiUrl":"https://doi.org/10.1038/s41598-025-34657-1","workflowStages":[]},"version":"v1","identity":"rs-7053119","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7053119","identity":"rs-7053119","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}