Artificial intelligence assisted panoramic radiography for enhanced caries diagnosis in clinical dental practice

Artificial intelligence assisted panoramic radiography for enhanced caries diagnosis in clinical dental practice | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Artificial intelligence assisted panoramic radiography for enhanced caries diagnosis in clinical dental practice Yujia Wu, Xiaowei Hou, Peng Ding, Zineng Xu, Hailong Bai, Lili Chen, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7234126/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Objectives This study aimed to evaluate the effectiveness of artificial intelligence (AI) panoramic radiography in improving caries detection accuracy in clinical dental practice, with a particular focus on enhancing diagnostic accuracy among early-career dentists. Materials and Methods In this retrospective diagnostic accuracy study, 15 dentists (12 young dentists with ≤ 3 years experience, 3 experienced dentists with > 10 years experience) interpreted 402 panoramic images under four conditions: AI-assisted, unassisted, AI-only, and expert reference. Primary endpoints included tooth-level sensitivity/specificity; secondary outcomes comprised interpretation time, case-level metrics, predictive values, and area under curve (AUC). Ground truth was established by three experienced experts using pixel-wise annotations. Results When assisted by the AI system, young dentists showed a marked improvement in sensitivity (Δ = 15.0%, 82.4% vs 67.4%, P < 0.001) without compromising specificity (97.4% vs 97.2%), suggesting that AI may help bridge the diagnostic experience gap. Interpretation time was also shortened by 22.7% (50.37 s vs 65.12 s, P = 0.003). Case-level analysis showed improved sensitivity (93.3% vs 84.7%, P < 0.001) and negative predictive value (93.9% vs 88.0%, P = 0.008). When operating independently, the AI system achieved 79.2% sensitivity and 98.4% specificity (AUC = 0.938), outperforming unassisted young dentists in both metrics ( P < 0.05). Conclusions AI-assisted panoramic radiography demonstrated clinically meaningful improvements in diagnostic performance and efficiency, particularly benefiting less experienced dentists. Dental Caries Radiography Panoramic Computer-Assisted Diagnosis Artificial Intelligence Figures Figure 1 Figure 2 Figure 3 Introduction Dental caries continues to be the most widespread chronic condition globally, with the 2022 WHO Global Oral Health Report estimating that 2.5 billion people live with untreated decay[ 1 ]. In China, the situation is especially pronounced—recent epidemiological studies indicate that over 88% of adults are affected, and the prevalence reaches nearly 98% among older adults[ 2 ]. Although early detection plays a critical role in enabling minimally invasive treatment, clinicians often face difficulties in diagnosing concealed lesions and achieving consistent interpretation between practitioners. In this context, panoramic radiography has gained traction as a practical screening modality. It provides a comprehensive view of the dentition while reducing radiation exposure by 30–50% compared to full-mouth intraoral series[ 3 , 4 ]. Research indicates that panoramic imaging, when employed for other diagnostic purposes, can reliably identify carious lesions as an incidental finding[ 5 ], and it enhances patient comfort by eliminating intraoral film placement. Its incidental caries detection capability supports initial assessments particularly in resource-constrained settings. However, significant diagnostic discrepancies persist, with general dentists missing 61% of carious lesions compared to specialists[ 6 ]. This highlights the need for tools to assist young dentists in accurately diagnosing caries from panoramic images. Artificial intelligence (AI) demonstrates transformative potential in caries detection, evolving from early mathematical models (Radon/DCT transformations)[ 7 ] to contemporary deep learning architectures. U-Net variants[ 8 ] and attention-enhanced networks[ 9 ] achieve 86–95% accuracy in controlled studies, yet clinical validation remains scarce. Notably, the sole clinically tested commercial system (Diagnocat) showed suboptimal sensitivity (44.5%) and reliability (ICC = 0.681)[ 10 ], underscoring the need for robust clinical trials bridging algorithmic development and real-world application. To address this gap, DeepCare (Beijing, China) developed a dual-stage U-Net system integrating tooth localization and caries detection. This multicenter trial evaluates whether AI assistance can: 1) Improve diagnostic sensitivity for young dentists (≤ 3 years experience); 2) Reduce interpretation time; 3) Achieve performance parity with experienced dentists (> 10 years experience). We hypothesize that AI augmentation will significantly enhance caries detection efficiency and accuracy compared to unaided panoramic image interpretation. Materials and methods Study Population and Image Acquisition A retrospective multicenter trial was conducted across three tertiary oral medical centers in China (Peking University School and Hospital of Stomatology, Tongji Medical College of Huazhong University of Science and Technology affiliated Tongji Hospital, and the Third Hospital of Hebei Medical University) from January to May 2022. Panoramic radiographs were acquired using five standardized devices: Carestream Dental CS 8000C (Carestream Dental, Atlanta, USA), KaVo OP 2D (KaVo, Biberach, Germany), NewTom VGi (Quantitative Radiology, Verona, Italy), Planmeca ProMax (Planmeca, Helsinki, Finland) and Sirona Orthophos XG (Dentsply Sirona, Bensheim, Germany). All images were de-identified to remove personal and institutional identifiers and stored in DICOM, JPEG, PNG, BMP, or TIFF formats. Inclusion criteria required patients ≥ 18 years old with diagnostic-quality images: full mandibular visibility without motion artifacts (signal-to-noise ratio ≥ 25 dB), resolution ≥ 800×1200 pixels (8-bit grayscale), and average grayscale values 70–180. Exclusion criteria included incomplete dental arches, full-crown restorations (> 50% crown coverage), residual roots only, and images used for AI training. The study protocol was approved by the Ethics Committee of Peking University School and Hospital of Stomatology (Approval No. PKUSSIRB-202277093a) with written informed consent obtained from all participants. AI System Development The AI system employed a dual-stage U-Net architecture, as illustrated in Fig. 1. Panoramic images typically capture the entire maxillofacial structure, making the caries areas appear relatively small. To address this, we previously trained a semantic segmentation network to automatically extract the region of interest (ROI), as described in our earlier work [ 11 ]. In that study, a U-Net-based algorithm was developed to segment the ROI around the alveolar crest, where the teeth are located, and was validated on a heterogeneous dataset of 6,046 panoramic radiographs, covering primary, mixed, and permanent dentitions. The algorithm demonstrated a high degree of accuracy, achieving a Dice coefficient of 0.963. This approach, leveraging the U-Net architecture, has shown outstanding performance in various medical image segmentation tasks[ 12 , 13 ]. The ROI segmentation results were used to crop the ROI image, which was further partitioned into image patches using a sliding window approach. These patches were then processed by a second U-Net model for caries segmentation. Given the similarity between caries and artifact patterns, caries segmentation presents a greater challenge than ROI segmentation. To enhance feature representation, we incorporated an attention mechanism into the decoder of U-Net. Specifically, a concurrent spatial and channel squeeze & excitation (scSE) module[ 14 ] was integrated after the convolution layers in each upsampling module to improve feature effectiveness in both spatial and channel dimensions. The framework was implemented in PyTorch[ 15 ], using the Adam optimizer[ 16 ] to minimize the loss functions. For ROI segmentation, we adopted the Dice coefficient loss function[ 17 ]. For caries segmentation, we used a combined loss[ 18 ]—a weighted sum of soft Dice loss and cross-entropy loss—to address the class imbalance issue (caries lesion pixels represent a small proportion). Both networks were trained with an initial learning rate of 0.0005 on an Nvidia RTX3090 GPU. Training convergence was determined by monitoring the validation loss; if the performance on the validation dataset remained unchanged for 20 epochs, the training process was stopped. Internal Validation of the AI System Prior to clinical validation, the caries-segmentation module was trained and internally validated using a dataset of 4,617 panoramic radiographs. The dataset was randomly split into training (70%), validation (10%), and testing (20%) subsets, corresponding to 3,231, 462, and 924 images, respectively. The validation set was used to fine-tune model hyperparameters. Final model performance was assessed on the internal test set using both pixel-level and lesion-level metrics. The model achieved a mean Dice coefficient of 0.764. Lesion-level sensitivity and specificity were 78.8% and 98.4%, respectively, based on a rule considering a predicted lesion correct if it overlapped with the expert-annotated ground truth. These results confirmed model robustness before proceeding to external clinical evaluation. Clinical Validation Workflow Reference Standard Establishment The reference standard was established by three experienced dentists with over 10 years of clinical experience, who performed pixel-wise annotations using a dedicated in-house tool under standardized conditions[ 19 ], without additional diagnostic aids or clinical records. Caries Definition and Severity Stratification Carious lesions were identified as radiolucent areas extending beyond the dentinoenamel junction on panoramic radiographs.Three board-certified endodontists (> 10 years experience) independently classified lesion depth according to radiographic penetration: Small caries: Radiolucency extends to the dentinoenamel junction or outer one-third of the dentin; Moderate caries: Radiolucency involving middle third of the dentin; Advanced caries: Radiolucency reaching inner third of the dentin or pulp proximity. The classification criteria were adapted from the American Dental Association Caries Classification System (ADA CCS)[ 20 ] and validated through pilot annotations. Final severity labels were assigned by majority vote, with discordant cases resolved via consensus discussion. This stratification enabled subgroup analyses to evaluate AI performance across disease progression stages. Study Design Twelve young dentists with less than 3 years of clinical experience independently evaluated the images both with and without AI assistance, following standardized training in image interpretation. The study flowchart is shown in Fig. 2. Images were categorized into four groups: experimental, control, AI, and standard. Evaluations were conducted in two sessions separated by a four-week washout period to minimize memory effects. Young dentists performed assessments in their routine clinical settings to simulate daily practice, recording the time required for each image evaluation (from image opening to completion and closure) while avoiding interruptions. All participants received detailed guidelines for image interpretation and annotation (Supplementary Material). The primary endpoints were the sensitivity and specificity of caries detection at the tooth level under AI-assisted and unassisted conditions. The secondary endpoints include interpretation time, sensitivity and specificity at the case level, positive predictive value, negative predictive value, and the AUC for caries detection at tooth and case levels. We also evaluated individual dentist performance to assess how AI assistance affected diagnostic accuracy at the personal level. Sample Size Calculation Based on preliminary results, we calculated the sample size to ensure sufficient power for detecting a 10% mean difference in sensitivity at the tooth level between AI-assisted and unassisted conditions. Assuming both readers and cases as random effects and accounting for variances and correlation coefficients from preliminary data, we planned to include 12 readers and at least 223 caries-positive cases. For specificity, to adequately demonstrate detection performance for negative cases, we performed a non-inferiority test with a margin of − 10% at the case level, resulting in a plan to include 12 readers and at least 96 caries-negative cases. Considering potential exclusions and to balance case distribution, we included 282 caries-positive and 120 caries-negative cases during initial screening. Statistical Analysis Methods We used the modified Obuchowski-Rockette model to analyze the primary endpoints of sensitivity and specificity at the tooth and case levels under AI-assisted and unassisted conditions. This model accounts for the hierarchical data structure, including random effects for readers and cases, and correlations among repeated measurements. The covariance matrix was estimated using the bootstrap method with 10,000 resampling, as specified in the statistical analysis plan. Based on this model, we calculated the differences and two-sided 95% confidence intervals (CIs) for sensitivity and specificity between AI-assisted and unassisted conditions. Superiority was established if the lower limit of the 95% CI for sensitivity was greater than 0%, and non-inferiority was established if the lower limit of the 95% CI for specificity was greater than − 10%. For individual dentist performance, we used paired t-tests to evaluate the impact of AI assistance on sensitivity and specificity improvements at the personal level. Differences between AI-assisted and unassisted conditions were calculated for each dentist, and confidence intervals were computed to assess statistical significance. Availability of Data and Code The inference model has been deployed as a secure web application ( https://cloud.deepcare.com/login ). Qualified researchers may request an academic trial account by contacting the corresponding author. Results Study Population A total of 402 subjects were retrospectively enrolled, consisting of 187 males and 215 females, with an average age of 39.47 ± 14.81 years. The cohort included 229 caries-positive cases (447 carious teeth) and 173 caries-negative cases (10,725 non-carious teeth). Among 447 carious teeth,1(0.2%) was small caries, 131(29.3%) was moderate caries and 315(70.5%) was advanced caries. At the case level, patients were classified by the most severe lesion present, yielding 44 moderate caries cases(19.2%) and 185 advanced caries cases(80.8%). Table 1 presents the comparative performance metrics at tooth and case levels. Reference Standard Reliability Inter-reader agreement among the three experienced dentists establishing the reference standard was excellent, with Fleiss Kappa = 0.82 (95% CI 0.79–0.85) at the tooth level. For young dentists, the average pairwise Kappa improved from 0.61 (95% CI 0.57–0.65) without AI to 0.73 (95% CI 0.69–0.77) with AI assistance. Tooth-Level Analysis AI assistance increased sensitivity by 15.0% (82.4% vs. 67.4%; Δ = 15.0%, 95% CI 9.8–20.2%, P < 0.001) while maintaining specificity (97.4% vs. 97.2%; Δ = 0.2%, P = 0.379). Positive predictive value (PPV) increased from 42.4–47.9% (Δ = 5.5%, 95% CI: 0.017–0.094, P = 0.005), suggesting that teeth identified as carious were more likely to be truly carious when AI was utilized. Negative predictive value (NPV) also increased from 99.1–99.5% (Δ = 0.4%, 95% CI: 0.002–0.006, P < 0.001), enhancing confidence in identifying non-carious teeth. The area under the curve (AUC) also increased from 0.825 to 0.901 with AI assistance (Δ = 7.6%, 95% CI: 0.038–0.114, P < 0.001), indicating an overall improvement in diagnostic accuracy. The standalone AI system achieved a sensitivity of 79.2%, specificity of 98.4%, PPV of 56.9%, and NPV of 99.4% at the tooth level, indicating high diagnostic capability as an adjunct tool. Stratified analysis revealed consistent improvements across caries severities: sensitivity increased by 21.1% for moderate caries (58.0%→79.1%, P < 0.001) and 14.1% for advanced caries (68.7%→82.8%, P = 0.002) (Table 2 ). Case-Level Analysis AI reduced interpretation time by 22.7% (50.37s vs. 65.12s; Δ=-14.75s, 95% CI -24.17 to -5.33, P = 0.003) while improving sensitivity (93.3% vs. 84.7%; Δ = 8.6%, P = 0.013). Specificity remained stable (59.8% vs. 61.0%; Δ=-1.2%, P = 0.774). The NPV increased significantly from 0.880 to 0.939 (Δ = 5.8%, 95% CI: 0.0157–0.1010, P = 0.008), enhancing confidence in ruling out caries-positive cases. The AUC also increased from 0.770 to 0.819 with AI assistance (difference of 0.049, 95% CI: 0.0046–0.0929, P = 0.031), indicating an overall enhancement in diagnostic performance. The standalone AI demonstrated a sensitivity of 92.1%, specificity of 63.0%, PPV of 59.4%, and NPV of 93.2% at the case level. AI assistance reduced missed diagnoses in patients with multiple carious teeth (≥ 2 teeth), improving full-caries detection rates by 21.7% (37.8%→59.5%, P < 0.001) . Performance in Challenging Cases In 62 subjects with non-carious dental defects (attrition/wedge defects), AI maintained specificity (95.8% vs. 95.6%, Δ=+0.2%, P = 0.412) while improving sensitivity (83.4% vs. 70.5%, Δ=+12.9%, P = 0.003), demonstrating robust discrimination against mimics (Table 3 ). Individual Performance All 12 young dentists showed sensitivity improvement with AI (mean Δ = 15.0%, range 5.1–37.4%, P < 0.001). Nine dentists exceeded standalone AI performance when assisted (Table 4 ). The most pronounced improvement occurred in Dentist #4 (sensitivity 75.2% vs. 37.8%; Δ = 37.4%). Specificity variations were nonsignificant (mean Δ = 0.2%, P = 0.452). Even among the three dentists whose sensitivity did not exceed that of the AI system, substantial improvements were noted, with gains of 10.52%, 14.54%, and 37.36%, respectively. Figure 3 illustrates the receiver operating characteristic (ROC) curves for each dentist without and with AI assistance, demonstrating the improvement in diagnostic performance when AI is utilized. The AUC increased from 0.825 without AI to 0.901 with AI at the tooth level (difference of 0.076, 95% CI: 0.038–0.114, P < 0.001), reflecting a significant improvement in overall diagnostic accuracy. Table 1 Comparative performance metrics of caries detection at the tooth level and case level Metric Tooth Level Case Level Unaided AI-Assisted Δ(95% CI) Unaided AI-Assisted Δ(95% CI) Sensitivity 67.4% 82.4% + 15.0%*(9.8–20.2) 84.7% 93.3% + 8.6%*(2.0-15.1) Specificity 97.2% 97.4% + 0.2%(-0.3-0.7) 61.0% 59.8% -1.2%(-9.2-6.8) PPV 42.4% 47.9% + 5.5%*(1.7–9.4) 57.3% 58.7% + 1.4%(-3.3-6.1) NPV 99.1% 99.5% + 0.4%*(0.2–0.6) 88.0% 93.9% + 5.9%*(1.6–10.1) AUC 0.825 0.901 + 0.076*(0.038–0.114) 0.770 0.819 + 0.049*(0.005–0.093) Time (s) - - 65.1 50.4 -14.7*(-24.2–5.3) P < 0.05; * P < 0.01; ** P < 0.001 Table 2 Stratified analysis by caries severity Severity Sensitivity (Δ%) 95% CI P Moderate + 21.1 11.4–30.8 < 0.001 Advanced + 14.1 7.4–20.8 0.002 Table 3 Performance in non-carious defects Metric Unaided AI-Assisted Δ (95% CI) Sensitivity 70.5% 83.4% + 12.9*(5.7–20.1) Specificity 95.6% 95.8% + 0.2 (-0.3-0.7) Table 4 Comparison of sensitivity and specificity for each dentist, both with and without AI assistance Dentist Sensitivity ΔSensitivity Specificity ΔSpecificity Unaided AI-Assisted Unaided AI-Assisted 1 81.0% 87.3% + 6.3% 96.5% 96.8% + 0.3% 2 66.4% 77.0% + 10.5% 98.7% 98.6% -0.1% 3 52.8% 82.8% + 30.0% 97.5% 97.4% -0.2% 4 37.8% 75.2% + 37.4% 99.3% 98.7% -0.6% 5 73.4% 85.7% + 12.3% 98.4% 98.5% + 0.2% 6 59.5% 83.2% + 23.7% 97.9% 97.4% -0.6% 7 76.3% 85.2% + 8.9% 97.3% 97.8% + 0.4% 8 75.8% 84.1% + 8.3% 96.3% 97.3% + 1.0% 9 74.1% 79.2% + 5.1% 92.3% 94.1% + 1.8% 10 74.1% 85.2% + 11.2% 97.6% 97.8% + 0.2% 11 76.7% 88.1% + 11.4% 96.7% 96.3% -0.4% 12 61.3% 75.8% + 14.5% 97.5% 98.0% + 0.6% Average 67.4% 82.4% + 15.0% 97.2% 97.4% + 0.2% Discussion The integration of artificial intelligence into dental diagnostics represents a transformative opportunity to address systemic inequities in oral healthcare delivery. In China, where dentist density has risen from 0.9 to 4.5 per 10,000 population over the past decade[ 21 ], persistent disparities in care quality remain pronounced between urban centers and under-resourced rural regions[ 22 , 23 ]. This technological advancement emerges as a strategic solution to bridge these geographical and socioeconomic divides. This multicenter trial establishes two pivotal advantages of AI-assisted panoramic radiography: 1) enhanced diagnostic reproducibility (inter-reader agreement κ improved from 0.61 to 0.73), and 2) reduced oversight in complex cases (21.7% sensitivity increase in multi-caries detection). These outcomes align with contemporary evidence demonstrating deep learning architectures like nnU-Net achieve caries staging accuracy comparable to experienced dentists (κ = 0.73 vs. 0.58) [ 24 ]. Particularly noteworthy is the system's differential performance across lesion severities - exhibiting 21.1% and 14.1% sensitivity gains for moderate and advanced caries respectively. Detecting small caries—particularly those limited to the enamel—remains inherently difficult, as early demineralization often fails to produce sufficient radiographic contrast. While our AI system demonstrated strong performance in identifying moderate and advanced caries, its ability to detect early-stage lesions is constrained by both image quality and the subtle nature of these changes. As such, despite the promise of AI as a diagnostic aid, accurate identification of small caries still depends on clinical examination and the integration of additional diagnostic tools.The clinical implications are magnified by China's caries epidemiology: national surveys reveal 80% adult prevalence with mean DMFT scores escalating from 4.54 (35-44y) to 13.33 (65-74y) [ 25 ]. In this context, the 95.8% specificity in discriminating non-carious defects (attrition/wedge lesions)[ 26 ], assumes particular significance, mitigating risks of overtreatment in resource-constrained settings where confirmatory imaging remains inaccessible. Operational benefits further underscore this technology's transformative potential. The observed 22.7% reduction in interpretation time enables scalable screening workflows. These gains mirror those reported in a recent Turkish study, where an AI model out-performed junior dentists (≤ 2 years’ experience), increasing sensitivity by 12–18 percentage points and almost halving reading time [ 27 ]. The concomitant rise in junior dentists’ diagnostic accuracy points to a practical strategy for redistributing expertise and narrowing the skills gap between novice and senior clinicians [ 28 ]. Early tele-dentistry implementations in low- and middle-income countries have achieved comparable efficiency gains (31% improvement in lesion detection) [ 29 ], further reinforcing the viability of this approach. Although the study produced encouraging results, several limitations should be acknowledged. First, in clinical settings, panoramic radiographs often lack the resolution and contrast needed for precise caries detection. These standard images may not capture sufficient detail for reliable diagnosis, limiting their utility when used in isolation. This highlights the practical challenge faced by dentists who must interpret panoramic radiographs without the aid of additional imaging modalities. Our study addresses this issue by demonstrating that an AI-assisted panoramic system can enhance diagnostic accuracy, highlighting its potential superiority in clinical applications. Secondly, the retrospective design of the study did not fully simulate the actual clinical operating environment. In practice, dentists may conduct intraoral examinations and take further radiographic images such as periapical or bite-wing radiographs to obtain more precise information for caries diagnosis, which was not included in this study. Third, although the reference standards were established by three experienced experts, their assessments may not reflect absolute diagnostic accuracy. Additionally, the limited number of participating clinical centers and their geographic distribution may constrain the generalizability of our findings. To strengthen real-world applicability, future prospective studies incorporating multimodal diagnostic approaches across broader regions will be essential. Furthermore, the establishment of reference standards through consensus panels rather than individual experts may enhance diagnostic benchmarking. AI-assisted systems have shown great promise in improving diagnostic accuracy and efficiency in dental radiography. However, we acknowledge concerns that reliance on AI could inadvertently lead to complacency in lesion detection. While AI systems can assist in identifying potential abnormalities, it is essential to recognize that they are designed to support—not replace—clinical judgment. Dentists must continue to examine radiographs critically, using their expertise to validate AI-generated outputs and ensure that subtle lesions are not missed. Although AI serves as a valuable tool in the diagnostic process, it cannot replicate the nuanced reasoning and contextual awareness that human clinicians contribute to patient care. Moreover, as AI systems—whether supervised or unsupervised—lack accountability, the ultimate responsibility for diagnosis and treatment rests with the clinician[ 30 ]. As such, dental education must evolve in parallel with technological advancements, ensuring that future practitioners are equipped to integrate AI tools thoughtfully and effectively into clinical practice. The WHO's mandate for equitable oral healthcare access finds tangible expression in this technology's deployment potential. By facilitating standardized caries surveillance across diverse clinical settings, AI-assisted screening has the potential to inform more targeted public health interventions—an increasingly urgent priority in light of China’s aging population and rising caries burden. Integrating this technology with telemedicine platforms and ongoing professional education initiatives may help drive systemic improvements in the delivery of preventive oral healthcare. Conclusion The AI-assisted panoramic radiography system demonstrated clinically meaningful improvements in diagnostic accuracy and efficiency, with the most significant benefits observed among early-career dentists. These results underscore the potential of artificial intelligence to support standardized caries screening and improve diagnostic consistency, particularly in clinical settings with limited resources. Declarations Compliance with Ethical Standards Conflict of Interest The authors declare no conflict of interest. Funding This work was supported by the Beijing Natural Science Foundation Haidian Original Innovation Joint Fund (Grant numbers [L242060]). Ethical Approval The study was approved by the Ethics Committee of Peking University School and Hospital of Stomatology (Approval No. PKUSSIRB-202277093a). A waiver of informed consent was granted due to the retrospective design and minimal risk involved. All patient data were anonymized to protect privacy. All procedures involving human participants were conducted in accordance with the ethical standards of the institutional and national research committee and with the 1964 Declaration of Helsinki and its later amendments. Informed Consent Informed consent was obtained from all individual participants included in the study. Study registration The study was conducted for product registration with the National Medical Products Administration and was not registered in a public database in order to protect proprietary commercial information. The clinical trial adhered to Good Clinical Practice (GCP) guidelines and was completed across three centers. The medical device clinical trial registration number is 20220252. Data entry and management were carried out by MedicalStrong, while data analysis and statistical analysis reports were completed by Peking Univerisity Chongqing Research Institute of Big Data, both of which are independent third-party organizations. References Organization WH. Global oral health status report: towards universal health coverage for oral health by 2030: executive summary, 2022. https://www.who.int/publications/i/item/9789240061484 . (Accessed May 5 2024). Wang X. Fourth National Oral Health Epidemiological Survey Report. 1 ed. Beijing: People's Medical Publishing House; 2018. Society of C, Endodontics CSA. [Guidelines for radiographic examination in cariology and endodontics]. Zhonghua Kou Qiang Yi Xue Za Zhi. 2021;56(4):311–7. Ludlow JB, Davies-Ludlow LE, White SC. Patient risk related to common dental radiographic examinations: the impact of 2007 International Commission on Radiological Protection recommendations regarding dose calculation. J Am Dent Assoc. 2008;139(9):1237–43. Wenzel A. Radiographic modalities for diagnosis of caries in a historical perspective: from film to machine-intelligence supported systems. Dentomaxillofac Radiol. 2021;50(5):20210010. Pakbaznejad EE, Pakkala T, Haukka J, Siukosaari P. Low reproducibility between oral radiologists and general dentists with regards to radiographic diagnosis of caries. Acta Odontol Scand. 2018;76(5):346–50. P S. S. P, Automated caries detection based on Radon transformation and DCT, 2017 8th International Conference on Computing, Communication and Networking Technologies (ICCCNT), 2017, pp. 1–6. Alharbi SS, AlRugaibah AA, Alhasson HF, Khan RU. Detection of Cavities from Dental Panoramic X-ray Images Using Nested U-Net Models. Appl Sci. 2023;13(23):12771. Zhu H, Cao Z, Lian L, Ye G, Gao H, Wu J. CariesNet: a deep learning approach for segmentation of multi-stage caries lesion from oral panoramic X-ray image. Neural Comput Appl (2022) 1–9. Mertens S, Krois J, Cantu AG, Arsiwala LT, Schwendicke F. Artificial intelligence for caries detection: Randomized trial. J Dent. 2021;115:103849. Xu M, Wu Y, Xu Z, Ding P, Bai H, Deng X. Robust automated teeth identification from dental radiographs using deep learning. J Dent 136 (2023). Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation, International Conference on Medical Image Computing and Computer-Assisted Intervention, 2015. Liu L, Cheng J, Quan Q, Wu FX, Wang J. A Survey on U-shaped networks in Medical Image Segmentations. Neurocomputing 409 (2020). Roy AG, Navab N, Wachinger C. Concurrent Spatial and Channel Squeeze & Excitation in Fully Convolutional Networks. Cham: Springer; 2018. Paszke A, Lerer A, Killeen T, Antiga L, Yang E, Tejani A, Fang L, Gross S, Bradbury J, Lin Z. PyTorch: An Imperative Style, High-Performance Deep Learning Library, Advances in Neural Information Processing Systems 32, Volume 11 of 20: 32nd Conference on Neural Information Processing Systems (NeurIPS 2019).Vancouver(CA).8–14 December 2019, 2020. Reddi SJ, Kale S, Kumar S. On the Convergence of Adam and Beyond, (2019). Sudre CH, Li W, Vercauteren T, Ourselin S, Cardoso MJ. Generalised Dice overlap as a deep learning loss function for highly unbalanced segmentations, International Workshop on Deep Learning in Medical Image Analysis International Workshop on Multimodal Learning for Clinical Decision Support, 2017. Taghanaki SA, Zheng Y, Zhou SK, Georgescu B, Sharma P, Xu D, Comaniciu D, Hamarneh G. Combo Loss: Handling Input and Output Imbalance in Multi-Organ Segmentation, (2018). Cantu AG, Gehrung S, Krois J, Chaurasia A, Rossi JG, Gaudin R, Elhennawy K, Schwendicke F. Detecting caries lesions of different radiographic extension on bitewings using deep learning. J Dent. 2020;100:103425. Young DA, Nový BB, Zeller GG, Hale R, Tran C. The American Dental Association Caries Classification System for clinical practice: a report of the American Dental Association Council on Scientific Affairs. J Am Dent Assoc. 2015;146(2):79–86. Organization WH. Dentists (per 10 000 population), 2024. https://www.who.int/data/gho/data/indicators/indicator-details/GHO/dentists-(per-10-000-population ). (Accessed May 29 2024). Li C, Yao NA. Socio-Economic Disparities in Dental Health and Dental Care Utilisation Among Older Chinese. Int Dent J. 2021;71(1):67–75. Qi X, Qu X, Wu B. Urban-Rural Disparities in Dental Services Utilization Among Adults in China's Megacities. Front Oral Health. 2021;2:673296. Lian L, Zhu T, Zhu F, Zhu H. Deep Learning for Caries Detection and Classification. Diagnostics (Basel) 11(9) (2021). Fourth N. Oral Health Epidemiologic Survey Report. Berghuis G, Cosyn J, De Bruyn H, Hommez G, Dierens M, Christiaens V. A controlled study on the diagnostic accuracy of panoramic and peri-apical radiography for detecting furcation involvement. BMC Oral Health. 2021;21(1):115. Gunec HG, Urkmez ES, Danaci A, Dilmac E, Onay HH, Cesur K, Aydin. Comparison of artificial intelligence vs. junior dentists' diagnostic performance based on caries and periapical infection detection on panoramic images. Quant Imaging Med Surg. 2023;13(11):7494–503. Liu P, Zhang X, Deng G, Guo W. Sociodemographic factors impacting the spatial distribution of private dental clinics in major cities of Peoples Republic of China. Int Dent J. 2024;74(5):1089–101. Ashtiani GH, Sabbagh S, Moradi S, Azimi S, Ravaghi V. Diagnostic accuracy of tele-dentistry in screening children for dental caries by community health workers in a lower‐middle‐income country. Int J Pediatr Dent. 2024;34(5):567–75. Shan T, Tay FR, Gu L. Application of Artificial Intelligence in Dentistry. J Dent Res. 2021;100(3):232–44. Supplementary Material Supplementary Material are not available with this version. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7234126","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":508329515,"identity":"7b8e7340-6aa2-42ee-924d-57ce6447d6c7","order_by":0,"name":"Yujia Wu","email":"","orcid":"","institution":"Peking University School and Hospital of Stomatology","correspondingAuthor":false,"prefix":"","firstName":"Yujia","middleName":"","lastName":"Wu","suffix":""},{"id":508329516,"identity":"527d80f7-f307-4727-9799-32e650e5e83a","order_by":1,"name":"Xiaowei Hou","email":"","orcid":"","institution":"Third Hospital of Hebei Medical University","correspondingAuthor":false,"prefix":"","firstName":"Xiaowei","middleName":"","lastName":"Hou","suffix":""},{"id":508329517,"identity":"49c3788b-390a-41a2-8ebc-c4fee66a3903","order_by":2,"name":"Peng Ding","email":"","orcid":"","institution":"deepcare Inc.","correspondingAuthor":false,"prefix":"","firstName":"Peng","middleName":"","lastName":"Ding","suffix":""},{"id":508329518,"identity":"285ff52e-1f27-430e-8b00-8c8841b3dd2c","order_by":3,"name":"Zineng Xu","email":"","orcid":"","institution":"deepcare Inc.","correspondingAuthor":false,"prefix":"","firstName":"Zineng","middleName":"","lastName":"Xu","suffix":""},{"id":508329520,"identity":"6d96d752-b878-4599-b9bc-712b4434d5cb","order_by":4,"name":"Hailong Bai","email":"","orcid":"","institution":"deepcare Inc.","correspondingAuthor":false,"prefix":"","firstName":"Hailong","middleName":"","lastName":"Bai","suffix":""},{"id":508329521,"identity":"6d770888-1bf9-4d1f-9bb3-f53aec9765fd","order_by":5,"name":"Lili Chen","email":"","orcid":"","institution":"Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Lili","middleName":"","lastName":"Chen","suffix":""},{"id":508329527,"identity":"9fb23642-9cb5-415b-93fc-8b55040a440e","order_by":6,"name":"Mingming Xu","email":"","orcid":"","institution":"Peking University School and Hospital of Stomatology","correspondingAuthor":false,"prefix":"","firstName":"Mingming","middleName":"","lastName":"Xu","suffix":""},{"id":508329528,"identity":"a3004307-22f5-4466-b032-6f683e0e501e","order_by":7,"name":"Xuliang Deng","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABF0lEQVRIiWNgGAWjYBACCQYGAzCDjxlEVhwAUww8xGhhA6s9Q5IWEMHYdoCBoBbJ9uSNjwt+HU5sY2d+9vDrvDvsujMSGB+8bWOQN8ehRZrnWbHxzL60xDZmNnNj2W3PmM1uJDAbzm1jMNzZgF2LnESOmTRvjw1QC4OZtOS2wyAtbNK8bQwJBgdwajH/zdsjAdTC/k1acg5YC/tvfFqkgbYw8/wA2cJjJvmxAWILMz4tkj3PiqV5G9KMgVrKpBmOAf1y5mGz5JxzEoYbcGiROJ688TPPn8Oy/fzHt0n+qLmTbHY8+eCHN2U28rhsYWBIAEUHhMkMjI5kILeBARxhOAFQC8MfCJPxBwODHR6lo2AUjIJRMEIBAB25WG+GhAIsAAAAAElFTkSuQmCC","orcid":"","institution":"Peking University School and Hospital of Stomatology","correspondingAuthor":true,"prefix":"","firstName":"Xuliang","middleName":"","lastName":"Deng","suffix":""}],"badges":[],"createdAt":"2025-07-28 12:53:37","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7234126/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7234126/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":90472517,"identity":"20b6f7bc-c716-4f0e-bc2d-bedcf18e5915","added_by":"auto","created_at":"2025-09-03 06:35:12","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":52109,"visible":true,"origin":"","legend":"\u003cp\u003eThe AI system employed a dual-stage U-Net architecture\u003c/p\u003e","description":"","filename":"OnlineFIGURE10318.png","url":"https://assets-eu.researchsquare.com/files/rs-7234126/v1/c92dd8b333c5671624a2d0a7.png"},{"id":90472520,"identity":"37139658-d04e-40b4-84b1-84de08264d33","added_by":"auto","created_at":"2025-09-03 06:35:12","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":74969,"visible":true,"origin":"","legend":"\u003cp\u003eThe study flowchart\u003c/p\u003e","description":"","filename":"OnlineFIGURE20318.png","url":"https://assets-eu.researchsquare.com/files/rs-7234126/v1/277af1db4f6bdd1e22fba8db.png"},{"id":90472518,"identity":"22d021a7-157b-4d03-87eb-6189b3276687","added_by":"auto","created_at":"2025-09-03 06:35:12","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":47805,"visible":true,"origin":"","legend":"\u003cp\u003eIllustrates the receiver operating characteristic (ROC) curves for each dentist without and with AI assistance, demonstrating the improvement in diagnostic performance when AI is utilized\u003c/p\u003e","description":"","filename":"FIGURE30318.png","url":"https://assets-eu.researchsquare.com/files/rs-7234126/v1/f12d13d98641d5958c3fe6c6.png"},{"id":94648632,"identity":"d7b631c2-ea42-4802-81a7-7d10264bc012","added_by":"auto","created_at":"2025-10-29 09:09:54","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1045975,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7234126/v1/be273ac0-327e-4ec5-bb06-6b9d046c38fd.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Artificial intelligence assisted panoramic radiography for enhanced caries diagnosis in clinical dental practice","fulltext":[{"header":"Introduction","content":"\u003cp\u003eDental caries continues to be the most widespread chronic condition globally, with the 2022 WHO Global Oral Health Report estimating that 2.5\u0026nbsp;billion people live with untreated decay[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. In China, the situation is especially pronounced\u0026mdash;recent epidemiological studies indicate that over 88% of adults are affected, and the prevalence reaches nearly 98% among older adults[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Although early detection plays a critical role in enabling minimally invasive treatment, clinicians often face difficulties in diagnosing concealed lesions and achieving consistent interpretation between practitioners.\u003c/p\u003e\u003cp\u003eIn this context, panoramic radiography has gained traction as a practical screening modality. It provides a comprehensive view of the dentition while reducing radiation exposure by 30\u0026ndash;50% compared to full-mouth intraoral series[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Research indicates that panoramic imaging, when employed for other diagnostic purposes, can reliably identify carious lesions as an incidental finding[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], and it enhances patient comfort by eliminating intraoral film placement. Its incidental caries detection capability supports initial assessments particularly in resource-constrained settings. However, significant diagnostic discrepancies persist, with general dentists missing 61% of carious lesions compared to specialists[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. This highlights the need for tools to assist young dentists in accurately diagnosing caries from panoramic images.\u003c/p\u003e\u003cp\u003eArtificial intelligence (AI) demonstrates transformative potential in caries detection, evolving from early mathematical models (Radon/DCT transformations)[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] to contemporary deep learning architectures. U-Net variants[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e] and attention-enhanced networks[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] achieve 86\u0026ndash;95% accuracy in controlled studies, yet clinical validation remains scarce. Notably, the sole clinically tested commercial system (Diagnocat) showed suboptimal sensitivity (44.5%) and reliability (ICC\u0026thinsp;=\u0026thinsp;0.681)[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], underscoring the need for robust clinical trials bridging algorithmic development and real-world application.\u003c/p\u003e\u003cp\u003eTo address this gap, DeepCare (Beijing, China) developed a dual-stage U-Net system integrating tooth localization and caries detection. This multicenter trial evaluates whether AI assistance can: 1) Improve diagnostic sensitivity for young dentists (\u0026le;\u0026thinsp;3 years experience); 2) Reduce interpretation time; 3) Achieve performance parity with experienced dentists (\u0026gt;\u0026thinsp;10 years experience). We hypothesize that AI augmentation will significantly enhance caries detection efficiency and accuracy compared to unaided panoramic image interpretation.\u003c/p\u003e"},{"header":"Materials and methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eStudy Population and Image Acquisition\u003c/h2\u003e\u003cp\u003e A retrospective multicenter trial was conducted across three tertiary oral medical centers in China (Peking University School and Hospital of Stomatology, Tongji Medical College of Huazhong University of Science and Technology affiliated Tongji Hospital, and the Third Hospital of Hebei Medical University) from January to May 2022. Panoramic radiographs were acquired using five standardized devices: Carestream Dental CS 8000C (Carestream Dental, Atlanta, USA), KaVo OP 2D (KaVo, Biberach, Germany), NewTom VGi (Quantitative Radiology, Verona, Italy), Planmeca ProMax (Planmeca, Helsinki, Finland) and Sirona Orthophos XG (Dentsply Sirona, Bensheim, Germany). All images were de-identified to remove personal and institutional identifiers and stored in DICOM, JPEG, PNG, BMP, or TIFF formats. Inclusion criteria required patients\u0026thinsp;\u0026ge;\u0026thinsp;18 years old with diagnostic-quality images: full mandibular visibility without motion artifacts (signal-to-noise ratio\u0026thinsp;\u0026ge;\u0026thinsp;25 dB), resolution\u0026thinsp;\u0026ge;\u0026thinsp;800\u0026times;1200 pixels (8-bit grayscale), and average grayscale values 70\u0026ndash;180. Exclusion criteria included incomplete dental arches, full-crown restorations (\u0026gt;\u0026thinsp;50% crown coverage), residual roots only, and images used for AI training. The study protocol was approved by the Ethics Committee of Peking University School and Hospital of Stomatology (Approval No. PKUSSIRB-202277093a) with written informed consent obtained from all participants.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eAI System Development\u003c/h3\u003e\n\u003cp\u003eThe AI system employed a dual-stage U-Net architecture, as illustrated in Fig.\u0026nbsp;1.\u003c/p\u003e\u003cp\u003ePanoramic images typically capture the entire maxillofacial structure, making the caries areas appear relatively small. To address this, we previously trained a semantic segmentation network to automatically extract the region of interest (ROI), as described in our earlier work [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. In that study, a U-Net-based algorithm was developed to segment the ROI around the alveolar crest, where the teeth are located, and was validated on a heterogeneous dataset of 6,046 panoramic radiographs, covering primary, mixed, and permanent dentitions. The algorithm demonstrated a high degree of accuracy, achieving a Dice coefficient of 0.963. This approach, leveraging the U-Net architecture, has shown outstanding performance in various medical image segmentation tasks[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThe ROI segmentation results were used to crop the ROI image, which was further partitioned into image patches using a sliding window approach. These patches were then processed by a second U-Net model for caries segmentation. Given the similarity between caries and artifact patterns, caries segmentation presents a greater challenge than ROI segmentation. To enhance feature representation, we incorporated an attention mechanism into the decoder of U-Net. Specifically, a concurrent spatial and channel squeeze \u0026amp; excitation (scSE) module[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] was integrated after the convolution layers in each upsampling module to improve feature effectiveness in both spatial and channel dimensions.\u003c/p\u003e\u003cp\u003eThe framework was implemented in PyTorch[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], using the Adam optimizer[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] to minimize the loss functions. For ROI segmentation, we adopted the Dice coefficient loss function[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. For caries segmentation, we used a combined loss[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]\u0026mdash;a weighted sum of soft Dice loss and cross-entropy loss\u0026mdash;to address the class imbalance issue (caries lesion pixels represent a small proportion). Both networks were trained with an initial learning rate of 0.0005 on an Nvidia RTX3090 GPU. Training convergence was determined by monitoring the validation loss; if the performance on the validation dataset remained unchanged for 20 epochs, the training process was stopped.\u003c/p\u003e\n\u003ch3\u003eInternal Validation of the AI System\u003c/h3\u003e\n\u003cp\u003ePrior to clinical validation, the caries-segmentation module was trained and internally validated using a dataset of 4,617 panoramic radiographs. The dataset was randomly split into training (70%), validation (10%), and testing (20%) subsets, corresponding to 3,231, 462, and 924 images, respectively. The validation set was used to fine-tune model hyperparameters. Final model performance was assessed on the internal test set using both pixel-level and lesion-level metrics. The model achieved a mean Dice coefficient of 0.764. Lesion-level sensitivity and specificity were 78.8% and 98.4%, respectively, based on a rule considering a predicted lesion correct if it overlapped with the expert-annotated ground truth. These results confirmed model robustness before proceeding to external clinical evaluation.\u003c/p\u003e\n\u003ch3\u003eClinical Validation Workflow\u003c/h3\u003e\n\u003cp\u003e\u003cb\u003eReference Standard Establishment\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe reference standard was established by three experienced dentists with over 10 years of clinical experience, who performed pixel-wise annotations using a dedicated in-house tool under standardized conditions[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e], without additional diagnostic aids or clinical records.\u003c/p\u003e\n\u003ch3\u003eCaries Definition and Severity Stratification\u003c/h3\u003e\n\u003cp\u003eCarious lesions were identified as radiolucent areas extending beyond the dentinoenamel junction on panoramic radiographs.Three board-certified endodontists (\u0026gt;\u0026thinsp;10 years experience) independently classified lesion depth according to radiographic penetration: Small caries: Radiolucency extends to the dentinoenamel junction or outer one-third of the dentin; Moderate caries: Radiolucency involving middle third of the dentin; Advanced caries: Radiolucency reaching inner third of the dentin or pulp proximity. The classification criteria were adapted from the American Dental Association Caries Classification System (ADA CCS)[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] and validated through pilot annotations. Final severity labels were assigned by majority vote, with discordant cases resolved via consensus discussion. This stratification enabled subgroup analyses to evaluate AI performance across disease progression stages.\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eStudy Design\u003c/h2\u003e\u003cp\u003eTwelve young dentists with less than 3 years of clinical experience independently evaluated the images both with and without AI assistance, following standardized training in image interpretation. The study flowchart is shown in Fig.\u0026nbsp;2. Images were categorized into four groups: experimental, control, AI, and standard. Evaluations were conducted in two sessions separated by a four-week washout period to minimize memory effects. Young dentists performed assessments in their routine clinical settings to simulate daily practice, recording the time required for each image evaluation (from image opening to completion and closure) while avoiding interruptions. All participants received detailed guidelines for image interpretation and annotation (Supplementary Material).\u003c/p\u003e\u003cp\u003eThe primary endpoints were the sensitivity and specificity of caries detection at the tooth level under AI-assisted and unassisted conditions. The secondary endpoints include interpretation time, sensitivity and specificity at the case level, positive predictive value, negative predictive value, and the AUC for caries detection at tooth and case levels. We also evaluated individual dentist performance to assess how AI assistance affected diagnostic accuracy at the personal level.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eSample Size Calculation\u003c/h3\u003e\n\u003cp\u003eBased on preliminary results, we calculated the sample size to ensure sufficient power for detecting a 10% mean difference in sensitivity at the tooth level between AI-assisted and unassisted conditions. Assuming both readers and cases as random effects and accounting for variances and correlation coefficients from preliminary data, we planned to include 12 readers and at least 223 caries-positive cases. For specificity, to adequately demonstrate detection performance for negative cases, we performed a non-inferiority test with a margin of \u0026minus;\u0026thinsp;10% at the case level, resulting in a plan to include 12 readers and at least 96 caries-negative cases. Considering potential exclusions and to balance case distribution, we included 282 caries-positive and 120 caries-negative cases during initial screening.\u003c/p\u003e\n\u003ch3\u003eStatistical Analysis Methods\u003c/h3\u003e\n\u003cp\u003eWe used the modified Obuchowski-Rockette model to analyze the primary endpoints of sensitivity and specificity at the tooth and case levels under AI-assisted and unassisted conditions. This model accounts for the hierarchical data structure, including random effects for readers and cases, and correlations among repeated measurements. The covariance matrix was estimated using the bootstrap method with 10,000 resampling, as specified in the statistical analysis plan. Based on this model, we calculated the differences and two-sided 95% confidence intervals (CIs) for sensitivity and specificity between AI-assisted and unassisted conditions. Superiority was established if the lower limit of the 95% CI for sensitivity was greater than 0%, and non-inferiority was established if the lower limit of the 95% CI for specificity was greater than \u0026minus;\u0026thinsp;10%.\u003c/p\u003e\u003cp\u003eFor individual dentist performance, we used paired t-tests to evaluate the impact of AI assistance on sensitivity and specificity improvements at the personal level. Differences between AI-assisted and unassisted conditions were calculated for each dentist, and confidence intervals were computed to assess statistical significance.\u003c/p\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003eAvailability of Data and Code\u003c/h2\u003e\u003cp\u003eThe inference model has been deployed as a secure web application (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://cloud.deepcare.com/login\u003c/span\u003e\u003cspan address=\"https://cloud.deepcare.com/login\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). Qualified researchers may request an academic trial account by contacting the corresponding author.\u003c/p\u003e\u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003eStudy Population\u003c/h2\u003e\u003cp\u003eA total of 402 subjects were retrospectively enrolled, consisting of 187 males and 215 females, with an average age of 39.47\u0026thinsp;\u0026plusmn;\u0026thinsp;14.81 years. The cohort included 229 caries-positive cases (447 carious teeth) and 173 caries-negative cases (10,725 non-carious teeth). Among 447 carious teeth,1(0.2%) was small caries, 131(29.3%) was moderate caries and 315(70.5%) was advanced caries. At the case level, patients were classified by the most severe lesion present, yielding 44 moderate caries cases(19.2%) and 185 advanced caries cases(80.8%). Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e presents the comparative performance metrics at tooth and case levels.\u003c/p\u003e\u003cp\u003e\u003cb\u003eReference Standard Reliability\u003c/b\u003e\u003c/p\u003e\u003cp\u003eInter-reader agreement among the three experienced dentists establishing the reference standard was excellent, with Fleiss Kappa\u0026thinsp;=\u0026thinsp;0.82 (95% CI 0.79\u0026ndash;0.85) at the tooth level. For young dentists, the average pairwise Kappa improved from 0.61 (95% CI 0.57\u0026ndash;0.65) without AI to 0.73 (95% CI 0.69\u0026ndash;0.77) with AI assistance.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003eTooth-Level Analysis\u003c/h2\u003e\u003cp\u003eAI assistance increased sensitivity by 15.0% (82.4% vs. 67.4%; Δ\u0026thinsp;=\u0026thinsp;15.0%, 95% CI 9.8\u0026ndash;20.2%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001) while maintaining specificity (97.4% vs. 97.2%; Δ\u0026thinsp;=\u0026thinsp;0.2%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.379).\u003c/p\u003e\u003cp\u003ePositive predictive value (PPV) increased from 42.4\u0026ndash;47.9% (Δ\u0026thinsp;=\u0026thinsp;5.5%, 95% CI: 0.017\u0026ndash;0.094, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.005), suggesting that teeth identified as carious were more likely to be truly carious when AI was utilized. Negative predictive value (NPV) also increased from 99.1\u0026ndash;99.5% (Δ\u0026thinsp;=\u0026thinsp;0.4%, 95% CI: 0.002\u0026ndash;0.006, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001), enhancing confidence in identifying non-carious teeth.\u003c/p\u003e\u003cp\u003eThe area under the curve (AUC) also increased from 0.825 to 0.901 with AI assistance (Δ\u0026thinsp;=\u0026thinsp;7.6%, 95% CI: 0.038\u0026ndash;0.114, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001), indicating an overall improvement in diagnostic accuracy.\u003c/p\u003e\u003cp\u003eThe standalone AI system achieved a sensitivity of 79.2%, specificity of 98.4%, PPV of 56.9%, and NPV of 99.4% at the tooth level, indicating high diagnostic capability as an adjunct tool.\u003c/p\u003e\u003cp\u003eStratified analysis revealed consistent improvements across caries severities: sensitivity increased by 21.1% for moderate caries (58.0%\u0026rarr;79.1%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and 14.1% for advanced caries (68.7%\u0026rarr;82.8%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.002) (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003eCase-Level Analysis\u003c/h2\u003e\u003cp\u003eAI reduced interpretation time by 22.7% (50.37s vs. 65.12s; Δ=-14.75s, 95% CI -24.17 to -5.33, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.003) while improving sensitivity (93.3% vs. 84.7%; Δ\u0026thinsp;=\u0026thinsp;8.6%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.013). Specificity remained stable (59.8% vs. 61.0%; Δ=-1.2%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.774).\u003c/p\u003e\u003cp\u003eThe NPV increased significantly from 0.880 to 0.939 (Δ\u0026thinsp;=\u0026thinsp;5.8%, 95% CI: 0.0157\u0026ndash;0.1010, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.008), enhancing confidence in ruling out caries-positive cases. The AUC also increased from 0.770 to 0.819 with AI assistance (difference of 0.049, 95% CI: 0.0046\u0026ndash;0.0929, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.031), indicating an overall enhancement in diagnostic performance.\u003c/p\u003e\u003cp\u003eThe standalone AI demonstrated a sensitivity of 92.1%, specificity of 63.0%, PPV of 59.4%, and NPV of 93.2% at the case level.\u003c/p\u003e\u003cp\u003eAI assistance reduced missed diagnoses in patients with multiple carious teeth (\u0026ge;\u0026thinsp;2 teeth), improving full-caries detection rates by 21.7% (37.8%\u0026rarr;59.5%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001) .\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003ePerformance in Challenging Cases\u003c/h2\u003e\u003cp\u003eIn 62 subjects with non-carious dental defects (attrition/wedge defects), AI maintained specificity (95.8% vs. 95.6%, Δ=+0.2%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.412) while improving sensitivity (83.4% vs. 70.5%, Δ=+12.9%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.003), demonstrating robust discrimination against mimics (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\u003ch2\u003eIndividual Performance\u003c/h2\u003e\u003cp\u003eAll 12 young dentists showed sensitivity improvement with AI (mean Δ\u0026thinsp;=\u0026thinsp;15.0%, range 5.1\u0026ndash;37.4%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Nine dentists exceeded standalone AI performance when assisted (Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). The most pronounced improvement occurred in Dentist #4 (sensitivity 75.2% vs. 37.8%; Δ\u0026thinsp;=\u0026thinsp;37.4%). Specificity variations were nonsignificant (mean Δ\u0026thinsp;=\u0026thinsp;0.2%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.452). Even among the three dentists whose sensitivity did not exceed that of the AI system, substantial improvements were noted, with gains of 10.52%, 14.54%, and 37.36%, respectively.\u003c/p\u003e\u003cp\u003eFigure 3 illustrates the receiver operating characteristic (ROC) curves for each dentist without and with AI assistance, demonstrating the improvement in diagnostic performance when AI is utilized. The AUC increased from 0.825 without AI to 0.901 with AI at the tooth level (difference of 0.076, 95% CI: 0.038\u0026ndash;0.114, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001), reflecting a significant improvement in overall diagnostic accuracy.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComparative performance metrics of caries detection at the tooth level and case level\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"7\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMetric\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e\u003cp\u003eTooth Level\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"3\" nameend=\"c7\" namest=\"c5\"\u003e\u003cp\u003eCase Level\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eUnaided\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAI-Assisted\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eΔ(95% CI)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eUnaided\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eAI-Assisted\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003eΔ(95% CI)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSensitivity\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e67.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e82.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;15.0%*(9.8\u0026ndash;20.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e84.7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e93.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e+\u0026thinsp;8.6%*(2.0-15.1)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e97.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e97.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;0.2%(-0.3-0.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e61.0%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e59.8%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e-1.2%(-9.2-6.8)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePPV\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e42.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e47.9%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;5.5%*(1.7\u0026ndash;9.4)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e57.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e58.7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e+\u0026thinsp;1.4%(-3.3-6.1)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNPV\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e99.1%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e99.5%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;0.4%*(0.2\u0026ndash;0.6)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e88.0%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e93.9%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e+\u0026thinsp;5.9%*(1.6\u0026ndash;10.1)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.825\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.901\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;0.076*(0.038\u0026ndash;0.114)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.770\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.819\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e+\u0026thinsp;0.049*(0.005\u0026ndash;0.093)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTime (s)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e- -\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e65.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e50.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e-14.7*(-24.2\u0026ndash;5.3)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"7\"\u003e\u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05; *\u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.01; **\u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eStratified analysis by caries severity\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSeverity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSensitivity (Δ%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003e95% CI\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u003cem\u003eP\u003c/em\u003e\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModerate\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e+\u0026thinsp;21.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e11.4\u0026ndash;30.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAdvanced\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e+\u0026thinsp;14.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e7.4\u0026ndash;20.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.002\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePerformance in non-carious defects\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMetric\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eUnaided\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAI-Assisted\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eΔ (95% CI)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSensitivity\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e70.5%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e83.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;12.9*(5.7\u0026ndash;20.1)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e95.6%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e95.8%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;0.2 (-0.3-0.7)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComparison of sensitivity and specificity for each dentist, both with and without AI assistance\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"7\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eDentist\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u003cp\u003eSensitivity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eΔSensitivity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eΔSpecificity\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eUnaided\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAI-Assisted\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eUnaided\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eAI-Assisted\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e81.0%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e87.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;6.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e96.5%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e96.8%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e+\u0026thinsp;0.3%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e66.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e77.0%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;10.5%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e98.7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e98.6%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e-0.1%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e52.8%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e82.8%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;30.0%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e97.5%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e97.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e-0.2%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e37.8%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e75.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;37.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e99.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e98.7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e-0.6%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e73.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e85.7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;12.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e98.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e98.5%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e+\u0026thinsp;0.2%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e59.5%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e83.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;23.7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e97.9%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e97.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e-0.6%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e76.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e85.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;8.9%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e97.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e97.8%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e+\u0026thinsp;0.4%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e75.8%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e84.1%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;8.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e96.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e97.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e+\u0026thinsp;1.0%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e74.1%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e79.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;5.1%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e92.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e94.1%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e+\u0026thinsp;1.8%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e74.1%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e85.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;11.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e97.6%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e97.8%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e+\u0026thinsp;0.2%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e76.7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e88.1%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;11.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e96.7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e96.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e-0.4%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e61.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e75.8%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;14.5%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e97.5%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e98.0%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e+\u0026thinsp;0.6%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAverage\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e67.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e82.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e+\u0026thinsp;15.0%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e97.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e97.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e+\u0026thinsp;0.2%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe integration of artificial intelligence into dental diagnostics represents a transformative opportunity to address systemic inequities in oral healthcare delivery. In China, where dentist density has risen from 0.9 to 4.5 per 10,000 population over the past decade[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], persistent disparities in care quality remain pronounced between urban centers and under-resourced rural regions[\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. This technological advancement emerges as a strategic solution to bridge these geographical and socioeconomic divides.\u003c/p\u003e\u003cp\u003eThis multicenter trial establishes two pivotal advantages of AI-assisted panoramic radiography: 1) enhanced diagnostic reproducibility (inter-reader agreement κ improved from 0.61 to 0.73), and 2) reduced oversight in complex cases (21.7% sensitivity increase in multi-caries detection). These outcomes align with contemporary evidence demonstrating deep learning architectures like nnU-Net achieve caries staging accuracy comparable to experienced dentists (κ\u0026thinsp;=\u0026thinsp;0.73 vs. 0.58) [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Particularly noteworthy is the system's differential performance across lesion severities - exhibiting 21.1% and 14.1% sensitivity gains for moderate and advanced caries respectively. Detecting small caries\u0026mdash;particularly those limited to the enamel\u0026mdash;remains inherently difficult, as early demineralization often fails to produce sufficient radiographic contrast. While our AI system demonstrated strong performance in identifying moderate and advanced caries, its ability to detect early-stage lesions is constrained by both image quality and the subtle nature of these changes. As such, despite the promise of AI as a diagnostic aid, accurate identification of small caries still depends on clinical examination and the integration of additional diagnostic tools.The clinical implications are magnified by China's caries epidemiology: national surveys reveal 80% adult prevalence with mean DMFT scores escalating from 4.54 (35-44y) to 13.33 (65-74y) [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. In this context, the 95.8% specificity in discriminating non-carious defects (attrition/wedge lesions)[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], assumes particular significance, mitigating risks of overtreatment in resource-constrained settings where confirmatory imaging remains inaccessible.\u003c/p\u003e\u003cp\u003eOperational benefits further underscore this technology's transformative potential. The observed 22.7% reduction in interpretation time enables scalable screening workflows. These gains mirror those reported in a recent Turkish study, where an AI model out-performed junior dentists (\u0026le;\u0026thinsp;2 years\u0026rsquo; experience), increasing sensitivity by 12\u0026ndash;18 percentage points and almost halving reading time [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. The concomitant rise in junior dentists\u0026rsquo; diagnostic accuracy points to a practical strategy for redistributing expertise and narrowing the skills gap between novice and senior clinicians [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. Early tele-dentistry implementations in low- and middle-income countries have achieved comparable efficiency gains (31% improvement in lesion detection) [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e], further reinforcing the viability of this approach.\u003c/p\u003e\u003cp\u003eAlthough the study produced encouraging results, several limitations should be acknowledged. First, in clinical settings, panoramic radiographs often lack the resolution and contrast needed for precise caries detection. These standard images may not capture sufficient detail for reliable diagnosis, limiting their utility when used in isolation. This highlights the practical challenge faced by dentists who must interpret panoramic radiographs without the aid of additional imaging modalities. Our study addresses this issue by demonstrating that an AI-assisted panoramic system can enhance diagnostic accuracy, highlighting its potential superiority in clinical applications. Secondly, the retrospective design of the study did not fully simulate the actual clinical operating environment. In practice, dentists may conduct intraoral examinations and take further radiographic images such as periapical or bite-wing radiographs to obtain more precise information for caries diagnosis, which was not included in this study. Third, although the reference standards were established by three experienced experts, their assessments may not reflect absolute diagnostic accuracy. Additionally, the limited number of participating clinical centers and their geographic distribution may constrain the generalizability of our findings. To strengthen real-world applicability, future prospective studies incorporating multimodal diagnostic approaches across broader regions will be essential. Furthermore, the establishment of reference standards through consensus panels rather than individual experts may enhance diagnostic benchmarking.\u003c/p\u003e\u003cp\u003eAI-assisted systems have shown great promise in improving diagnostic accuracy and efficiency in dental radiography. However, we acknowledge concerns that reliance on AI could inadvertently lead to complacency in lesion detection. While AI systems can assist in identifying potential abnormalities, it is essential to recognize that they are designed to support\u0026mdash;not replace\u0026mdash;clinical judgment. Dentists must continue to examine radiographs critically, using their expertise to validate AI-generated outputs and ensure that subtle lesions are not missed. Although AI serves as a valuable tool in the diagnostic process, it cannot replicate the nuanced reasoning and contextual awareness that human clinicians contribute to patient care. Moreover, as AI systems\u0026mdash;whether supervised or unsupervised\u0026mdash;lack accountability, the ultimate responsibility for diagnosis and treatment rests with the clinician[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. As such, dental education must evolve in parallel with technological advancements, ensuring that future practitioners are equipped to integrate AI tools thoughtfully and effectively into clinical practice.\u003c/p\u003e\u003cp\u003eThe WHO's mandate for equitable oral healthcare access finds tangible expression in this technology's deployment potential. By facilitating standardized caries surveillance across diverse clinical settings, AI-assisted screening has the potential to inform more targeted public health interventions\u0026mdash;an increasingly urgent priority in light of China\u0026rsquo;s aging population and rising caries burden. Integrating this technology with telemedicine platforms and ongoing professional education initiatives may help drive systemic improvements in the delivery of preventive oral healthcare.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThe AI-assisted panoramic radiography system demonstrated clinically meaningful improvements in diagnostic accuracy and efficiency, with the most significant benefits observed among early-career dentists. These results underscore the potential of artificial intelligence to support standardized caries screening and improve diagnostic consistency, particularly in clinical settings with limited resources.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eCompliance with Ethical Standards\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflict of Interest\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no conflict of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the Beijing Natural Science Foundation Haidian Original Innovation Joint Fund (Grant numbers [L242060]).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical Approval\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study was approved by the Ethics Committee of Peking University School and Hospital of Stomatology (Approval No. PKUSSIRB-202277093a). A waiver of informed consent was granted due to the retrospective design and minimal risk involved. All patient data were anonymized to protect privacy. All procedures involving human participants were conducted in accordance with the ethical standards of the institutional and national research committee and with the 1964 Declaration of Helsinki and its later amendments.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInformed Consent\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eInformed consent was obtained from all individual participants included in the study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStudy registration\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study was conducted for product registration with the National Medical Products Administration and was not registered in a public database in order to protect proprietary commercial information. The clinical trial adhered to Good Clinical Practice (GCP) guidelines and was completed across three centers. The medical device clinical trial registration number is 20220252. Data entry and management were carried out by MedicalStrong, while data analysis and statistical analysis reports were completed by Peking Univerisity Chongqing Research Institute of Big Data, both of which are independent third-party organizations.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eOrganization WH. Global oral health status report: towards universal health coverage for oral health by 2030: executive summary, 2022. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.who.int/publications/i/item/9789240061484\u003c/span\u003e\u003cspan address=\"https://www.who.int/publications/i/item/9789240061484\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (Accessed May 5 2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang X. Fourth National Oral Health Epidemiological Survey Report. 1 ed. Beijing: People's Medical Publishing House; 2018.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSociety of C, Endodontics CSA. [Guidelines for radiographic examination in cariology and endodontics]. Zhonghua Kou Qiang Yi Xue Za Zhi. 2021;56(4):311\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLudlow JB, Davies-Ludlow LE, White SC. Patient risk related to common dental radiographic examinations: the impact of 2007 International Commission on Radiological Protection recommendations regarding dose calculation. J Am Dent Assoc. 2008;139(9):1237\u0026ndash;43.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWenzel A. Radiographic modalities for diagnosis of caries in a historical perspective: from film to machine-intelligence supported systems. Dentomaxillofac Radiol. 2021;50(5):20210010.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePakbaznejad EE, Pakkala T, Haukka J, Siukosaari P. Low reproducibility between oral radiologists and general dentists with regards to radiographic diagnosis of caries. Acta Odontol Scand. 2018;76(5):346\u0026ndash;50.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eP S. S. P, Automated caries detection based on Radon transformation and DCT, 2017 8th International Conference on Computing, Communication and Networking Technologies (ICCCNT), 2017, pp. 1\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAlharbi SS, AlRugaibah AA, Alhasson HF, Khan RU. Detection of Cavities from Dental Panoramic X-ray Images Using Nested U-Net Models. Appl Sci. 2023;13(23):12771.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhu H, Cao Z, Lian L, Ye G, Gao H, Wu J. CariesNet: a deep learning approach for segmentation of multi-stage caries lesion from oral panoramic X-ray image. Neural Comput Appl (2022) 1\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMertens S, Krois J, Cantu AG, Arsiwala LT, Schwendicke F. Artificial intelligence for caries detection: Randomized trial. J Dent. 2021;115:103849.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXu M, Wu Y, Xu Z, Ding P, Bai H, Deng X. Robust automated teeth identification from dental radiographs using deep learning. J Dent 136 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRonneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation, International Conference on Medical Image Computing and Computer-Assisted Intervention, 2015.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu L, Cheng J, Quan Q, Wu FX, Wang J. A Survey on U-shaped networks in Medical Image Segmentations. Neurocomputing 409 (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRoy AG, Navab N, Wachinger C. Concurrent Spatial and Channel Squeeze \u0026amp; Excitation in Fully Convolutional Networks. Cham: Springer; 2018.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePaszke A, Lerer A, Killeen T, Antiga L, Yang E, Tejani A, Fang L, Gross S, Bradbury J, Lin Z. PyTorch: An Imperative Style, High-Performance Deep Learning Library, Advances in Neural Information Processing Systems 32, Volume 11 of 20: 32nd Conference on Neural Information Processing Systems (NeurIPS 2019).Vancouver(CA).8\u0026ndash;14 December 2019, 2020.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eReddi SJ, Kale S, Kumar S. On the Convergence of Adam and Beyond, (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSudre CH, Li W, Vercauteren T, Ourselin S, Cardoso MJ. Generalised Dice overlap as a deep learning loss function for highly unbalanced segmentations, International Workshop on Deep Learning in Medical Image Analysis International Workshop on Multimodal Learning for Clinical Decision Support, 2017.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTaghanaki SA, Zheng Y, Zhou SK, Georgescu B, Sharma P, Xu D, Comaniciu D, Hamarneh G. Combo Loss: Handling Input and Output Imbalance in Multi-Organ Segmentation, (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCantu AG, Gehrung S, Krois J, Chaurasia A, Rossi JG, Gaudin R, Elhennawy K, Schwendicke F. Detecting caries lesions of different radiographic extension on bitewings using deep learning. J Dent. 2020;100:103425.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYoung DA, Nov\u0026yacute; BB, Zeller GG, Hale R, Tran C. The American Dental Association Caries Classification System for clinical practice: a report of the American Dental Association Council on Scientific Affairs. J Am Dent Assoc. 2015;146(2):79\u0026ndash;86.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOrganization WH. Dentists (per 10 000 population), 2024. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.who.int/data/gho/data/indicators/indicator-details/GHO/dentists-(per-10-000-population\u003c/span\u003e\u003cspan address=\"https://www.who.int/data/gho/data/indicators/indicator-details/GHO/dentists-(per-10-000-population\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). (Accessed May 29 2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi C, Yao NA. Socio-Economic Disparities in Dental Health and Dental Care Utilisation Among Older Chinese. Int Dent J. 2021;71(1):67\u0026ndash;75.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eQi X, Qu X, Wu B. Urban-Rural Disparities in Dental Services Utilization Among Adults in China's Megacities. Front Oral Health. 2021;2:673296.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLian L, Zhu T, Zhu F, Zhu H. Deep Learning for Caries Detection and Classification. Diagnostics (Basel) 11(9) (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFourth N. Oral Health Epidemiologic Survey Report.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBerghuis G, Cosyn J, De Bruyn H, Hommez G, Dierens M, Christiaens V. A controlled study on the diagnostic accuracy of panoramic and peri-apical radiography for detecting furcation involvement. BMC Oral Health. 2021;21(1):115.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGunec HG, Urkmez ES, Danaci A, Dilmac E, Onay HH, Cesur K, Aydin. Comparison of artificial intelligence vs. junior dentists' diagnostic performance based on caries and periapical infection detection on panoramic images. Quant Imaging Med Surg. 2023;13(11):7494\u0026ndash;503.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu P, Zhang X, Deng G, Guo W. Sociodemographic factors impacting the spatial distribution of private dental clinics in major cities of Peoples Republic of China. Int Dent J. 2024;74(5):1089\u0026ndash;101.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAshtiani GH, Sabbagh S, Moradi S, Azimi S, Ravaghi V. Diagnostic accuracy of tele-dentistry in screening children for dental caries by community health workers in a lower‐middle‐income country. Int J Pediatr Dent. 2024;34(5):567\u0026ndash;75.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShan T, Tay FR, Gu L. Application of Artificial Intelligence in Dentistry. J Dent Res. 2021;100(3):232\u0026ndash;44.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Supplementary Material","content":"\u003cp\u003eSupplementary Material are not available with this version.\u003c/p\u003e\n"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Dental Caries, Radiography, Panoramic, Computer-Assisted Diagnosis, Artificial Intelligence","lastPublishedDoi":"10.21203/rs.3.rs-7234126/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7234126/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eObjectives\u003c/h2\u003e\u003cp\u003eThis study aimed to evaluate the effectiveness of artificial intelligence (AI) panoramic radiography in improving caries detection accuracy in clinical dental practice, with a particular focus on enhancing diagnostic accuracy among early-career dentists.\u003c/p\u003e\u003ch2\u003eMaterials and Methods\u003c/h2\u003e\u003cp\u003eIn this retrospective diagnostic accuracy study, 15 dentists (12 young dentists with \u0026le;\u0026thinsp;3 years experience, 3 experienced dentists with \u0026gt;\u0026thinsp;10 years experience) interpreted 402 panoramic images under four conditions: AI-assisted, unassisted, AI-only, and expert reference. Primary endpoints included tooth-level sensitivity/specificity; secondary outcomes comprised interpretation time, case-level metrics, predictive values, and area under curve (AUC). Ground truth was established by three experienced experts using pixel-wise annotations.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eWhen assisted by the AI system, young dentists showed a marked improvement in sensitivity (Δ\u0026thinsp;=\u0026thinsp;15.0%, 82.4% vs 67.4%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001) without compromising specificity (97.4% vs 97.2%), suggesting that AI may help bridge the diagnostic experience gap. Interpretation time was also shortened by 22.7% (50.37 s vs 65.12 s, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.003). Case-level analysis showed improved sensitivity (93.3% vs 84.7%, \u003cem\u003eP\u0026thinsp;\u0026lt;\u003c/em\u003e\u0026thinsp;0.001) and negative predictive value (93.9% vs 88.0%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.008). When operating independently, the AI system achieved 79.2% sensitivity and 98.4% specificity (AUC\u0026thinsp;=\u0026thinsp;0.938), outperforming unassisted young dentists in both metrics (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05).\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e\u003cp\u003eAI-assisted panoramic radiography demonstrated clinically meaningful improvements in diagnostic performance and efficiency, particularly benefiting less experienced dentists.\u003c/p\u003e","manuscriptTitle":"Artificial intelligence assisted panoramic radiography for enhanced caries diagnosis in clinical dental practice","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-03 06:35:07","doi":"10.21203/rs.3.rs-7234126/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"3981986b-31de-4c40-961c-a363f43fcc8d","owner":[],"postedDate":"September 3rd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-10-29T09:09:21+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-03 06:35:07","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7234126","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7234126","identity":"rs-7234126","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}