Deep learning model development and clinical validation for radiographic surrogate markers of implant esthetic risk Authors: Hengyi Liu, Zhuohong Gong, Beichen Wen, Longshiyu Qiu, Xiaofei Meng, Gengbin Cai, Peisheng Zeng, Shijie Chen, Mengru Shi, Xinchun Zhang, Zhuofan Chen, Ruixuan Wang, Zetao Chen Published: 2026 ABSTRACT AbstractReliable risk assessment for implant-supported restorations in the esthetic zone is critical yet challenging due to complex anatomical variations and the inherent subjectivity of traditional clinical assessments. To address these limitations, we developed a multi-functional artificial intelligence (AI) system designed to automate the assessment of radiographic surrogate markers of implant esthetic risk, specifically periapical inflammation (INFLAM), adjacent tooth restorations (RESTOR), and the distance between the contact point and alveolar crest (DISTAN) from periapical radiographs. The system underwent rigorous validation through a four-pronged strategy: direct performance comparison against dentists of varying experience, a human-AI collaboration scenario, exploratory prospective clinical testing and multi-site validation. Results demonstrated that the AI matched junior dentists in INFLAM/RESTOR tasks while statistically outperforming experts in the DISTAN task. Crucially, human-AI integration revealed a task-specific synergistic effect, particularly in DISTAN assessments, where it significantly enhanced recall compared to isolated performance. Furthermore, an exploratory prospective clinical testing and multi-site validation demonstrated the system’s consistent performance and acceptable generalization ability, achieving high specificity across diverse clinical settings. This versatile AI tool facilitates the precise, objective assessment of radiographic surrogate markers strongly associated with esthetic risk. Although direct clinical esthetic outcomes were not prospectively measured, the system’s proven ability to enhance dentist performance highlights its promising potential for pre-implant evaluation and clinical decision support.Subject terms: Computational biology and bioinformatics, Health care, Medical research Introduction A harmonious esthetic appearance is a vital component of dental health and quality of life1. Nonetheless, it can be compromised by a variety of factors, including pathological conditions such as tooth defects, dentition abnormalities, and traumatic injuries. These disruptions not only affect physical appearance but can also lead to significant functional impairments. Reconstruction of these esthetic deficiencies, including using cosmetic dentistry and reconstructive surgical interventions to restore the harmonious relationship of soft and hard tissues, not only improves health but also meets the growing demands for functional and esthetic rehabilitation2,3. The key to successful dental esthetic implant treatment lies in the ability to accurately evaluate pre-surgical risk factors, so that dentists can select the most appropriate intervention strategies. However, assessing the radiographic determinants associated with esthetic risk presents considerable challenges due to substantial interindividual anatomical variability, patient-specific preference and iatrogenic factors4. For instance, implant-supported restorations, as an important method for restoring missing teeth, evaluating the radiographic surrogate markers for their final esthetic reconstruction, especially the anterior teeth region (commonly known as “esthetic zone”), is highly complex and influenced by multiple variables. The dental implant-supported esthetic outcome is influenced by patient factors (e.g., morphology of remaining soft and hard tissue), iatrogenic factors (e.g., techniques of the implantologist, symmetry between the restoration and adjacent teeth), and patient-specific factors (e.g., personalized perception of beauty and satisfaction). To facilitate the reliable assessment of radiographic surrogate markers of implant esthetic risk, various evaluation methods have been proposed. Among them, periapical radiography remains one of the most accessible, cost-effective, and widely used tools in dental clinical practice. In the context of radiographic surrogate markers evaluation, these radiographs provide crucial esthetic-determining information about the edentulous site, adjacent teeth, and surrounding anatomical structures. More specifically, periapical pathology of adjacent teeth can cause destruction of the buccal bone plate, leading to mid-facial mucosal recession and gingival instability (INFLAM)5–9; Furthermore, the presence of restorations on adjacent teeth complicates optical harmony and increases the risk of interproximal bone resorption due to biomechanical failure (RESTOR)10,11; Finally, an excessive vertical distance from the contact point to the alveolar crest (≥7 mm) drastically reduces the probability of papilla fill, indicating a high risk for the formation of “black triangles”(DISTAN)12,13 (Fig. 1). The aforementioned factors play a pivotal role in the successful risk assessment of the results of dental implant-supported esthetic reconstruction. Evaluation of surrogate markers of implant esthetic risk using periapical radiographs. Three key risk factors were manually evaluated: infection in adjacent teeth (INFLAM; favorable if no low-density shadow >1 mm at implant or apical sites, suboptimal if present), existing restorations in adjacent teeth (RESTOR; favorable if absent, suboptimal if high-density restorations present), and the distance from the contact point to the alveolar crest (DISTAN; favorable if <7 mm, suboptimal if ≥7 mm). The combination of surrogate markers determined the preoperative esthetic risk assessment. a Case one: Favorable manual risk assessments for INFLAM, RESTOR, and DISTAN, consistent with minimal postoperative esthetic concerns. b Case two: Favorable risk assessments for INFLAM and RESTOR, suboptimal for DISTAN, corresponding to a postoperative interdental “black triangle” between tooth #31 and tooth #41. Evaluating radiographic determinants of dental implant-supported esthetic risk via these surrogate markers remains subjective and heavily relies on dentists’ experience, posing significant challenges to standardized, reliable, and reproducible risk assessment14–16. With the rapid development of artificial intelligence (AI), this cutting-edge technology has been used in dental crown design, color shading and implant surgery planning. However, AI-based radiographic surrogate markers risk assessment through periapical radiography still faces various challenges in the scenario of dentistry. Firstly, periapical radiography of tooth-missing patients poses a greater challenge in the absence of key landmarks, making tooth and root identification more difficult, increasing the likelihood of misclassification17. Besides, current AI models primarily focus on crown detection, often overlooking endodontically treated teeth and resin fillings with a high chance of discoloration, which complicates accurate classification due to highly heterogeneous radiographic characteristics18. Furthermore, quantitative assessment of distances between the contact point and the alveolar bone crest adds another layer of complexity, demanding high levels of precision that are rarely achieved using periapical radiography19. Hence, it is necessary to develop integrated AI architectures and explore novel application strategies to support automated risk assessment of esthetic outcomes for implant dentistry20. Despite numerous studies confirming the strength of AI in preclinical sessions, the actual integration of AI-based dental implant-supported esthetic-related radiographic surrogate markers assessment tools into routine clinical workflows remains highly challenging21,22. The inherent subjectivity, professional complexity, and precision required in esthetic evaluations can lead to discrepancies between AI-based radiographic surrogate markers' assessment of implant esthetic risk and human clinical judgment, raising concerns about reliability and adoption23. Furthermore, previous AI models were trained on retrospective datasets and lack validation in real-world environments24. This underscores the need for prospective clinical trials based on real-world data to evaluate the generalization ability, accuracy, and usability of these tools in diverse patient populations. The aim of this study was to address these challenges by developing and validating a versatile AI system for assessing key radiographic surrogate markers influencing esthetic outcomes (INFLAM, RESTOR, and DISTAN) in the anterior esthetic zone. The specific objectives were: (1) To construct a multi-functional AI system for the simultaneous detection, classification, and segmentation of these factors from periapical radiographs; (2) To evaluate the system’s performance against dental experts and in a human-AI integration scenario; and (3) To validate its clinical utility and generalization ability through both an exploratory prospective clinical test and a multi-site external validation. We hypothesized that this AI tool would not only achieve high accuracy but also enhance the diagnostic performance of dentists, thereby improving the reliability of evaluating radiographic surrogate markers associated with esthetic risk prior to dental implant-supported reconstruction. Results Baseline characteristics This study incorporates two sequential stages: initial system development using a retrospective dataset, followed by clinical validation involving exploratory prospective and multi-site datasets. The retrospective dataset consisted of periapical radiographs from 318 patients, comprising 161 males and 157 females, with an average age of 41.1 ± 15.3 years. The exploratory prospective test and multi-site validation dataset consisted of 49 (29 male and 20 female) and 27 (14 male and 13 female) radiographs, respectively. The radiographic surrogate markers of implant esthetic risk were classified into favorable and suboptimal. The proportions of suboptimal samples identified for the three tasks were as follows: 36.8%, 40.8% and 33.3% for the INFLAM task, 39.3%, 53.1% and 55.6% for the RESTOR task and 61.0%, 51.0% and 63.0% for the DISTAN task on the retrospective, exploratory prospective and multi-site validation datasets, respectively. The detailed demographic characteristics and the distribution of suboptimal samples for the three tasks (INFLAM, RESTOR, and DISTAN) are presented in Table 1. To ensure consistency, intra-rater repeatability of the ground truth annotation was assessed on a subset of data, yielding Kappa values of 0.945 (INFLAM), 0.962 (RESTOR), and 0.920 (DISTAN), confirming robust data quality. The inter-rater agreement was also excellent, with a Cohen’s Kappa of 0.913 for the INFLAM task, 0.980 for the RESTOR task, and 0.894 for the DISTAN task. Table 1Demographic characteristics and ground truth sample distribution for the retrospective, prospective test, and multi-site validation datasetsParameterNumberRetrospective dataExploratory prospective test dataMulti-site validation dataDemographic characteristicsSex Female1572013 Male1612914Age (Years) Mean Age41.1 ± 15.336.3 ± 13.446.7 ± 17.8 Range18.1–85.018.8–69.319.0–82.6Distribution of sample for each taskINFLAM task (Ground truth) Favorable outcome2012918 Suboptimal outcome117209RESTOR task (Ground truth) Favorable outcome1932312 Suboptimal outcome1252615DISTAN task (Ground truth) Favorable outcome1112410 Suboptimal outcome1942517 Excluded1300Patient age was calculated based on the date of birth (accurate to the month) and the date of the radiographic imaging. In the DISTAN task, the 13 samples labeled “Excluded” were cases where the crown of the adjacent tooth was too severely damaged, precluding reliable manual measurement. Demographic characteristics and ground truth sample distribution for the retrospective, prospective test, and multi-site validation datasets Patient age was calculated based on the date of birth (accurate to the month) and the date of the radiographic imaging. In the DISTAN task, the 13 samples labeled “Excluded” were cases where the crown of the adjacent tooth was too severely damaged, precluding reliable manual measurement. System performance on the validation set The detection model demonstrated no false detections or missed detections in the INFLAM & RESTOR task, establishing it as a reliable foundation for region of interest (ROI) extraction. The detection model with a classifier achieved the highest average accuracies of 0.827 and 0.952 in the INFLAM and RESTOR tasks, respectively. In the DISTAN task, the segmentation model achieved a mean Dice score of 0.921. The segmentation masks were further processed using the Douglas-Peucker algorithm to approximate an ideal quadrilateral, from which key points were extracted to estimate the distance value. System performance on the internal test set The proposed AI system demonstrated robust performance across all three tasks, showcasing acceptable generalization ability on unseen data. Overall, the system achieved high receiver operating characteristic curve (AUROC) values for the INFLAM (0.921, 95%CI: 0.864–0.967) and RESTOR (0.942, 95%CI: 0.874–0.989) tasks, indicating strong classification ability. For the challenging DISTAN task, the system achieved a high F1-score (0.911, 95%CI: 0.851–0.946) and precision (0.973, 95%CI: 0.904–0.992) (Fig. 2). These results collectively demonstrate the effectiveness of the AI system in integrating object detection, segmentation, and classification in this study (Fig. 3A). Performance of the AI system in evaluating radiographic surrogate markers of implant esthetic risk for INFLAM, RESTOR, and DISTAN tasks in the internal test set. For INFLAM, it showed high accuracy and robust discriminative ability, with precision surpassing 0.77 and AUROC exceeding 0.92. The RESTOR task achieved even higher accuracy above 0.92, with balanced precision and recall and an AUROC close to 0.94. The DISTAN task delivered particularly precise segmentation, combining high accuracy with precision approaching 0.97 and strong recall, supported by AUROC and area under the precision-recall curve (AUPRC) values around 0.94 and 0.90, reflecting consistent and reliable performance. Performance of the AI system across the three different validation cohorts. The rows represent the three validation cohorts: A Internal Test Set; B Exploratory Prospective Validation Set; C Multi-Site Validation Set. The columns represent the three assessment tasks: (Left) INFLAM Task; (Center) RESTOR Task; (Right) DISTAN Task. For the matrix labels, “Y” represents suboptimal, and “N” represents favorable. To explicitly address the inherent challenge of dichotomizing a continuous anatomical measurement, a post-hoc subgroup analysis was conducted to compare the system’s performance on borderline cases versus non-borderline cases relative to the 7 mm clinical threshold. When evaluating cases strictly within the borderline [6.0, 8.0] mm range, the accuracy decreased to 63.16%. In an even narrower boundary of [6.5, 7.5] mm, the accuracy was 62.50%, with a specificity of 75.00% and a sensitivity of 50.00%. For all cases falling clearly outside the boundary region (continuous distance < 6.0 mm or >8.0 mm), the AI system achieved reliable agreement with the human-derived ground truth, yielding 100% accuracy, 100% sensitivity, and 100% specificity. Crucially, these results demonstrate that 100% of the misclassifications in the DISTAN task were tightly clustered around the artificial 7 mm threshold. Comparison and integration between dentists and an AI system The comparison of dental implant-supported esthetic-related radiographic surrogate markers assessment between the AI system and dentists showcased the system’s comparative advantages. Figure 4 presents the confusion matrices for dentists of varying experience levels. Specifically, the AI system achieved an accuracy of 77.05% in the INFLAM task (Table 2), which is comparable to the performance of junior dentists (70.49%, p = 0.244), and the system’s performance was slightly lower than that of senior (85.25%, p = 0.101). However, the system’s performance is significantly lower than expert dentists (87.70%, p = 0.029). In the RESTOR task (Table 3), the AI system’s accuracy was 90.16%, which was just slightly lower than that of senior (92.62%, p = 0.607) and expert (93.44%, p = 0.388) dentists, but higher than that of junior dentists (88.52%, p = 0.832). However, none of these differences were found to be statistically significant. In the DISTAN task (Table 4), the AI system’s accuracy was 88.52%, which is statistically significantly higher (p < 0.001 vs. junior and senior; p = 0.009 vs. expert) than all other groups (Fig. 5A). Performance of dentists of different experience levels in evaluating radiographic surrogate markers of implant esthetic risk. The rows represent the three tasks: A INFLAM Task; B RESTOR Task; C DISTAN Task. The columns represent the dentist’s experience level: (Left) Expert; (Center) Senior; (Right) Junior. For the matrix labels, “Y” represents suboptimal, and “N” represents favorable. The matrices show that dentist performance varies significantly by both task and experience. Table 2Performance of the AI system and dentists on the INFLAM taskGroupAccuracy (95% CI)Precision (95% CI)Specificity (95% CI)F1-score (95% CI)Recall (95% CI)AI Alone77.05% (68.83–83.62%)75.00% (61.22–85.08%)82.86% (72.38–89.91%)72.00% (61.22–81.42%)69.23% (55.73–80.09%)Junior70.49% (61.87–77.86%)80.77% (62.12–91.49%)92.86% (84.34–96.91%)53.85% (39.34–67.44%)40.38% (28.16–53.93%)Junior + AI77.87% (69.72–84.32%)71.93% (59.17–81.92%)77.14% (66.05–85.41%)a75.23% (64.76–83.72%)a78.85% (65.97–87.76%)aSenior85.25% (77.88–90.46%)90.48% (77.93–96.23%)94.29% (86.21–97.76%)80.85% (71.42–88.89%)73.08% (59.75–83.23%)Senior + AI81.97% (74.20–87.78%)75.00% (62.77–84.22%)a78.57% (67.61–86.56%)a80.36% (71.43–87.93%)b86.54% (74.73–93.32%)a,bExpert87.70% (80.70–92.41%)100.00% (90.59–100.00%)100.00% (94.80–100.00%)83.15% (73.97–90.91%)71.15% (57.73–81.67%)Expert + AI83.61% (76.03–89.13%)b78.57% (66.18–87.29%)a,b82.86% (72.38–89.91%)a81.48% (72.89–88.89%)b84.62% (72.48–91.99%)a,bBoth statistical comparisons were based on McNemar’s test with a Bonferroni-adjusted threshold (Bonferroni-adjusted threshold, p < 0.0167).adenotes a statistically significant difference between the integrated “Dentist + AI” modality and the dentist working alone.bdenotes a statistically significant difference between the integrated modality and the AI system operating alone. Performance of the AI system and dentists on the INFLAM task Both statistical comparisons were based on McNemar’s test with a Bonferroni-adjusted threshold (Bonferroni-adjusted threshold, p < 0.0167). adenotes a statistically significant difference between the integrated “Dentist + AI” modality and the dentist working alone. bdenotes a statistically significant difference between the integrated modality and the AI system operating alone. Table 3Performance of the AI system and dentists on the RESTOR taskGroupAccuracy (95% CI)Precision (95% CI)Specificity (95% CI)F1-score (95% CI)Recall (95% CI)AI Alone90.16% (80.16–95.41%)80.95% (60.00–92.33%)90.48% (77.93–96.23%)85.00% (70.96–95.66%)89.47% (68.61–97.06%)Junior88.52% (81.66–93.04%)77.27% (63.01–87.16%)88.10% (79.46–93.40%)82.93% (72.22–91.36%)89.47% (75.87–95.83%)Junior + AI86.89% (79.75–91.76%)70.37% (57.17–80.86%)b80.95% (71.28–87.92%)b82.61% (72.94–89.89%)100.00% (90.82–100.00%)Senior92.62% (86.57–96.07%)93.94% (80.39–98.32%)97.62% (91.73–99.34%)87.32% (77.14–94.74%)81.58% (66.58–90.78%)Senior + AI92.62% (86.57–96.07%)82.22% (68.67–90.71%)a90.48% (82.32–95.09%)89.16% (81.39–95.15%)97.37% (86.51–99.53%)Expert93.44% (87.59–96.64%)91.67% (78.17–97.13%)96.43% (90.02–98.78%)89.19% (81.25–95.77%)86.84% (72.67–94.25%)Expert + AI90.98% (84.57–94.89%)80.00% (66.18–89.10%)a89.29% (80.88–94.26%)86.75% (78.38–93.62%)94.74% (82.71–98.54%)Both statistical comparisons were based on McNemar’s test with a Bonferroni-adjusted threshold (Bonferroni-adjusted threshold, p < 0.0167).adenotes a statistically significant difference between the integrated “Dentist + AI” modality and the dentist working alone.bdenotes a statistically significant difference between the integrated modality and the AI system operating alone. Performance of the AI system and dentists on the RESTOR task Both statistical comparisons were based on McNemar’s test with a Bonferroni-adjusted threshold (Bonferroni-adjusted threshold, p < 0.0167). adenotes a statistically significant difference between the integrated “Dentist + AI” modality and the dentist working alone. bdenotes a statistically significant difference between the integrated modality and the AI system operating alone. Table 4Performance of the AI system and dentists on the DISTAN taskGroupAccuracy (95% CI)Precision (95% CI)Specificity (95% CI)F1-score (95% CI)Recall (95% CI)AI Alone88.52% (78.16–94.33%)97.30% (86.18–99.52%)94.74% (75.36–99.06%)91.14% (83.78–96.63%)85.71% (72.16–93.28%)Junior59.84% (50.97–68.11%)70.11% (59.81–78.72%)31.58% (19.08–47.46%)71.35% (62.79–78.92%)72.62% (62.25–81.01%)Junior + AI76.23% (67.95–82.91%)a,b75.70% (66.78–82.84%)a,b31.58% (19.08–47.46%)b84.82% (78.92–89.66%)a96.43% (90.02–98.78%)a,bSenior53.28% (44.46–61.90%)69.01% (57.52–78.56%)42.11% (27.85–57.81%)63.23% (53.50–72.29%)58.33% (47.65–68.29%)Senior + AI78.69% (70.60–85.02%)a78.43% (69.50–85.30%)a,b42.11% (27.85–57.81%)b86.02% (80.23–90.72%)a95.24% (88.39–98.13%)a,bExpert75.41% (67.07–82.20%)86.49% (76.88–92.49%)73.68% (57.99–85.03%)81.01% (73.82–86.79%)76.19% (66.06–84.03%)Expert + AI87.70% (80.70–92.41%)a87.91% (79.64–93.11%)b71.05% (55.24–83.00%)b91.43% (86.59–95.30%)a95.24% (88.39–98.13%)a,bBoth statistical comparisons were based on McNemar’s test with a Bonferroni-adjusted threshold (Bonferroni-adjusted threshold, p < 0.0167).adenotes a statistically significant difference between the integrated “Dentist + AI” modality and the dentist working alone.bdenotes a statistically significant difference between the integrated modality and the AI system operating alone. Performance of the AI system and dentists on the DISTAN task Both statistical comparisons were based on McNemar’s test with a Bonferroni-adjusted threshold (Bonferroni-adjusted threshold, p < 0.0167). adenotes a statistically significant difference between the integrated “Dentist + AI” modality and the dentist working alone. bdenotes a statistically significant difference between the integrated modality and the AI system operating alone. Comparative and integrated performance of the AI system and dentists in evaluating radiographic surrogate markers of implant esthetic risk. A–C Comparisons between the AI system and dentists of different levels in evaluating radiographic surrogate markers of implant esthetic risk. A INFLAM Task; B RESTOR task; C DISTAN task. The AI system matched junior dentists in the INFLAM task, was slightly lower than senior and expert dentists in the RESTOR task, and outperformed all groups in the DISTAN task. D–F Integrations of AI system and dentists of different levels in evaluating radiographic surrogate markers of implant esthetic risk. D INFLAM Task; E RESTOR task; F DISTAN task. The combination improved recall in INFLAM and RESTOR tasks, significantly enhanced dentists’ performance in the DISTAN task. For the integration between dentists and the AI system, the results varied significantly by task, as detailed in the confusion matrices in Fig. 6. For the RESTOR task (Table 3), the integration showed no statistically significant difference in accuracy compared to the AI system alone or any of the dentist groups. For the INFLAM task (Table 2), the integrated system, particularly when used by senior dentists (p = 0.016), achieved a higher recall than the AI system operating alone. In the DISTAN task (Table 4), the combination of the AI system and dental practitioners resulted in statistically significantly higher (p = 0.003 vs. expert; p < 0.001 vs. junior and senior) accuracy compared to all dentist groups working alone. (Fig. 5B). Integrations of the AI system and dentists of different experience levels on the risk assessment of radiographic surrogate markers. The rows represent the three tasks: A INFLAM Task; B RESTOR Task; C DISTAN Task. The columns represent the integration partner: (Left) AI + Expert; (Center) AI + Senior; (Right) AI + Junior. For the matrix labels, “Y” represents suboptimal, and “N” represents favorable. The matrices show a task-dependent integration effect. Performance of the AI system on the exploratory prospective clinical validation test The study evaluated the performance of the system across three tasks on the exploratory prospective clinical validation test (Fig. 3B). The INFLAM task demonstrated an accuracy of 83.67%, sensitivity of 95.00%, specificity of 75.86%, positive predictive value (PPV) of 73.08%, and negative predictive value (NPV) of 95.65%. RESTOR task exhibited highly reliable classification with an accuracy of 97.96%, sensitivity of 96.15%, specificity of 100%, PPV of 100%, and NPV of 95.83%. DISTAN task achieved an accuracy of 85.71%, sensitivity of 100%, specificity of 70.83%, PPV of 78.13%, and NPV of 100%. While the RESTOR task showed the best overall performance, the DISTAN task demonstrated slightly lower specificity (Table 5). Table 5Performance AI system in evaluating radiographic surrogate markers of implant esthetic risk for INFLAM, RESOTR and DISTAN tasks in the prospective test setTaskAccuracy (95% CI)Sensitivity (95% CI)Specificity (95% CI)PPV (95% CI)NPV (95% CI)INFLAM83.67% (70.34%–92.68%)95.00% (75.13%–99.87%)75.86% (56.46%–89.70%)73.08% (52.21%–88.43%)95.65% (78.05%–99.89%)RESTOR97.96% (89.15%–99.95%)96.15% (80.36%–99.90%)100.00% (85.18%–100.00%)100.00% (86.28%–100.00%)95.83% (78.88%–99.89%)DISTAN85.71% (72.76–94.06%)100.00% (86.28–100.00%)70.83% (48.91–87.38%)78.13% (60.03–90.72%)100.00% (80.49–100.00%)For INFLAM, it achieved a solid accuracy of 0.8367. The RESTOR task demonstrated highly reliable classification, reaching almost 0.98 accuracy, sensitivity exceeding 0.96, good specificity, and PPV. The DISTAN task achieved the similar overall accuracy as INFLAM, with higher sensitivity but lower specificity. Performance AI system in evaluating radiographic surrogate markers of implant esthetic risk for INFLAM, RESOTR and DISTAN tasks in the prospective test set For INFLAM, it achieved a solid accuracy of 0.8367. The RESTOR task demonstrated highly reliable classification, reaching almost 0.98 accuracy, sensitivity exceeding 0.96, good specificity, and PPV. The DISTAN task achieved the similar overall accuracy as INFLAM, with higher sensitivity but lower specificity. System performance on the multi-site validation The study further evaluated the system’s generalization ability on the pooled multi-site validation data. Overall, the system demonstrated consistent performance across these unseen data (Fig. 3C). The INFLAM task achieved an excellent and balanced performance, with a good accuracy of 88.89% and strong sensitivity. The RESTOR task continued to show the highest performance, achieving a good accuracy of 96.30%, which was characterized by high specificity and high positive predictive value. For the DISTAN task, the system achieved a strong accuracy of 88.89%, with good specificity and PPV as well. These external validation results support the system’s potential for robust deployment across different clinical settings (Table 6). Table 6Performance AI system in evaluating radiographic surrogate markers of implant esthetic risk for INFLAM, RESTOR, and DISTAN tasks in the multi-site validation setTaskAccuracy (95% CI)Sensitivity (95% CI)Specificity (95% CI)PPV (95% CI)NPV (95% CI)INFLAM88.89% (70.84%–97.65%)88.89% (51.75%–99.72%)88.89% (65.29%–98.62%)80.00% (44.39%–97.48%)94.12% (71.31%–99.85%)RESTOR96.30% (81.03%–99.91%)93.33% (68.05–99.83%)100.00% (73.54–100.00%)100.00% (76.84–100.00%)92.31% (63.97–99.81%)DISTAN88.89% (70.84–97.65%)82.35% (56.57–96.20%)100.00% (69.15–100.00%)100.00% (76.84–100.00%)76.92% (46.19–94.96%)For INFLAM, it achieved a high accuracy of 0.89, with well-balanced sensitivity and specificity. The RESTOR task demonstrated good classification, reaching 0.96 accuracy, high sensitivity, and good specificity and PPV. The DISTAN task achieved a good accuracy of 0.89, and also demonstrated high specificity and PPV, albeit with a more moderate sensitivity. Performance AI system in evaluating radiographic surrogate markers of implant esthetic risk for INFLAM, RESTOR, and DISTAN tasks in the multi-site validation set For INFLAM, it achieved a high accuracy of 0.89, with well-balanced sensitivity and specificity. The RESTOR task demonstrated good classification, reaching 0.96 accuracy, high sensitivity, and good specificity and PPV. The DISTAN task achieved a good accuracy of 0.89, and also demonstrated high specificity and PPV, albeit with a more moderate sensitivity. Discussion This study presents an AI-based solution for the assessment of radiographic surrogate markers influencing dental implant-supported esthetic outcome. The system subsequently underwent rigorous exploratory validation, including both multi-site and prospective testing, demonstrating acceptable generalization ability and clinically satisfactory performance. More importantly, these results highlight two central findings that define the system’s clinical value: (1) its ability to perform specific objective tasks (e.g., the DISTAN task) at a level statistically superior to human experts, and (2) its role as a complementary partner, where human-AI integration achieved a synergistic effect, yielding higher accuracies than either the AI or the dentists working alone. These findings suggest that a well-trained AI system serves as a valuable assistant for clinical decision support, capable of improving the reliability of esthetic risk assessment in dental implantology. Addressing the complex nature of assessment of radiographic surrogate markers of implant esthetic risk requires tailored solutions for specific clinical challenges. To accurately identify edentulous spaces, high-resolution image segmentation can overcome inaccuracies caused by alveolar resorption. For apical inflammation, our study leverages a hybrid network architecture that combines EfficientDet and DenseNet. This two-stage approach first uses EfficientDet for robust detection of potential areas, which are then precisely classified by DenseNet. In these inflammation-focused tasks, our system’s performance reached the upper threshold of junior-to-senior-level clinical proficiency. Recent research demonstrates the potential of AI in detecting periapical lesions. For instance, studies by Li et al.25 and Gunec et al.26 have shown high F1-scores and sensitivities for periapical periodontitis. However, many existing studies focus on general periapical disease of an affected tooth, whereas our INFLAM task is specifically designed to assess inflammation in the teeth adjacent to the potential implant site, which poses a different diagnostic challenge. Unlike general practice involving acute symptoms, the priority in implantology is excluding latent or asymptomatic risks that compromise the alveolar bone bed. Therefore, our system targeted objective radiographic signs (Periapical Index (PAI) ≥ 3) to detect these “silent” anomalies, acting as a critical safety net. Given the challenging requirements of the screening context, this study employed ROC-based optimization to select the optimized operating point guided by Youden’s Index statistics27. At this optimized threshold, the INFLAM model achieved a recall of 70.00% and an accuracy of 88.46%. The model’s high AUROC of 92.10%, along with the improved recall, confirms the system’s discriminative ability. This optimization strategy enhances the AI system’s utility as a screening tool, enabling it to more effectively identify cases at high esthetic risk and demonstrating substantial clinical value. Notably, in the context of our “OR” integration strategy, combining a high-recall (though lower-specificity) AI with human evaluation yields more comprehensive performance across both recall and specificity. This aligns with the primary clinical objective for AI-assisted screening, that is, to establish a robust, high-sensitivity safety net that minimizes missed cases in diverse clinical scenarios. Similar to the INFLAM task in this study, a hybrid network was also employed for the RESTOR task, combining EfficientDet for detection with DenseNet for classification. In this task, the system performs at a proficiency level comparable to the upper range of a junior-to-mid-level expert. Recent studies have confirmed the promise of deep learning (using models like Faster R-CNN or YOLO) in detecting dental restorations, often achieving high precision28,29. It has been noted that this system can perform comparably to or even better than dentists with several years of experience, though perhaps not surpassing senior specialists. Our findings align with this literature: for restoration-related tasks, current AI models, including ours, operate effectively at an upper-junior to mid-level of clinical proficiency, highlighting their potential to serve as powerful assistive tools that enhance clinical workflows and standardize patient care. Measuring distances requires pixel-level accuracy to quantify spatial relationships or dimensions, making it a segmentation task. The nnU-Net is designed explicitly for high-precision medical image segmentation and is adept at handling the semi-quantitative analysis required for accurate distance measurements30. For the DISTAN task, it operates with over expert-level precision and consistency, effectively automating high-fidelity analyses in prosthetic detection. However, despite the high precision of nnU-Net in segmentation and distance measurement, it is important to acknowledge that radiographic imaging itself inherently contains systematic errors, such as centering discrepancies between different devices, which can affect measurement accuracy31. Therefore, a margin of error within ±1 mm of the threshold is clinically acceptable. When applying this AI tool in practice, special attention should be given to cases near the threshold, where errors are more likely. Human-AI collaboration is essential to reduce misclassification and ensure reliable clinical outcomes. While many studies have explored caries and restoration detection using deep learning32–35, little attention has been given to the distance between the contact point and the alveolar ridge, particularly in terms of measuring this distance. This measurement is critical for assessing implant surgery and restoration success, as it directly evaluates bone support and ensures the long-term stability of implants. Traditionally, this manual measurement is subjective and prone to errors. In this study, our automated method significantly outperformed all human groups, including experts, by reliably identifying landmarks and objectively calculating this distance. This capability directly reduces human error and minimizes the risk of complications such as “black triangles”. The study by Le et al.36 on cephalometric landmark detection and our experiment on dental implant esthetics both highlight the value of human-AI collaboration, though they reveal different performance dynamics. The AI model alone was superior to inexperienced dental students, achieving a successful detection rate of 73.17% compared to the students’ 47.40%. Their collaboration modestly improved the beginners’ accuracy by 5.33%, but this combined effort did not surpass the AI’s standalone performance. In contrast, our experiment revealed a more complex relationship, where the AI’s accuracy varied, sometimes surpassing and sometimes falling short of dentists’, depending on the specific task and the dentists’ experience level. Crucially, our results demonstrated a more powerful synergy; the dentist-AI integration on the DISTAN task produced accuracies much higher than the dentists achieved alone. This indicates that while AI support is beneficial to dentists, the collaboration in our implant risk assessment tasks created a superior hybrid intelligence, outperforming any single contributor, a key difference from the cephalometric study, where the AI remained the unchallenged top performer. Meanwhile, in a detailed comparison with human dentists, the AI system demonstrated tiered performance: its accuracy was comparable to junior dentists for the INFLAM task and showed no significant difference from any dentist group for the RESTOR task. Most notably, the system exhibited excellent capability in the DISTAN task, where its accuracy was statistically and significantly higher than that of all human practitioners, including seasoned experts. Our study further evaluated the performance of the system across three tasks on the exploratory prospective clinical validation test. While the RESTOR task showed the best overall performance, the DISTAN task’s lower specificity indicates room for further refinement. Notably, our multi-site validation also achieved results that exceeded the training-level performance. Our system’s clinical utility was rigorously validated through a multi-faceted evaluation framework. First, human-AI comparison demonstrated its non-inferiority to high-tier specialists in accuracy and efficiency, confirming its capacity to replicate expert-level clinical reconstruction risk assessment for targeted applications. Second, human-AI integration trials revealed the optimal assistive role of the technology: it functions as a synergistic partner that augments junior dentists by reducing predictive variability, enhances senior practitioners’ throughput by automating quantitative analyses, and enables domain experts to prioritize complex case management, thus establishing a tiered augmentation paradigm aligned with clinical hierarchies. A critical aspect of this human-AI integration strategy is the inherent diagnostic trade-off introduced by the “OR” consensus rule. By mathematically pooling the positive assessments of both the human and the AI, this approach was specifically designed to function as a rigorous clinical “safety net”. As explicitly demonstrated in our results (Tables 2–4), while this strategy did not uniformly yield statistically significant improvements in overall accuracy, particularly in the INFLAM and RESTOR tasks when subjected to stringent Bonferroni corrections, it systematically and significantly prioritized Recall. For instance, the integration propelled Recall to highly reliable levels (>95%) in the anatomically complex DISTAN task across all practitioner levels. In the context of pre-surgical radiographic risk assessment, this trade-off is highly desirable. We prioritize a highly sensitive screening baseline that minimizes irreversible surgical complications by preventing false negatives, accepting that the resulting slight decrease in specificity can be prudently managed through subsequent clinical verification. Third, an exploratory prospective clinical evaluation study and multi-site validation confirmed the system’s operational robustness in real-world settings, demonstrating consistent performance across various patient demographics, imaging equipment variations, and unstructured clinical environments. In case-based validation, the system demonstrated critical prognostic utility (Fig. 7): Case One exhibited favorable reconstruction outcome flags for inflammation and existing restorations but suboptimal distance analysis, accurately highlighting the postoperative interdental “black triangle” complication despite achieving chromatic harmony with adjacent teeth. Conversely, Case Two triggered suboptimal predictive outcomes alerts for both INFLAM and RESTOR tasks, foreshadowing clinically significant outcomes, including perceptible crown discoloration and apical periodontitis in neighboring teeth requiring long-term monitoring. However, the dentist made the new ceramic crowns instead of the old porcelain-fused metal crowns to keep the color consistent. These results validate the system’s capacity to predict both dental implant-supported esthetic reconstruction and functional prognosis, underscoring its role in preemptive treatment planning and patient counseling. Case-based validation of the AI system for evaluating radiographic surrogate markers of implant esthetic risk. a Case one: Favorable AI risk assessments for INFLAM and RESTOR, suboptimal for DISTAN, matching the postoperative interdental “black triangle” between tooth #11 and tooth #21. b Case two: Suboptimal AI risk assessments for INFLAM and RESTOR, favorable for DISTAN, corresponding to postoperative crown discoloration and apical periodontitis at tooth #21. The clinical implications of this system are substantial, primarily because its robust validation addresses a critical gap where many other studies remain confined to single-site datasets. It was confirmed in both an exploratory prospective clinical test and a multi-site validation cohort, the latter of which demonstrated consistent reliability, such as high specificity on RESTOR and DISTAN tasks. This real-world value is rooted in a dual capacity that defines its clinical role. The system demonstrated an ability to outperform human experts in objective, qualitative tasks, as its accuracy on the DISTAN task was statistically superior to all clinician groups. Furthermore, it acted as a true complementary partner in more subjective assessments. Unlike other human-AI collaboration study36, where the AI’s standalone performance remained superior, our integration analysis revealed a more powerful synergy on the INFLAM and RESTOR tasks, where the combined team’s accuracy was higher than either the AI or the clinician working alone. This dual capacity to outperform human judgment in objective measurements and synergistically complement it in subjective assessments establishes the system as a decision-support tool. By providing objective and quantitative assistance, it can help clinicians to identify sub-clinical risks and guide interventions toward more predictable and successful esthetic outcomes. Despite the promising results, this study has several limitations. The first limitation concerns the data and the generalization ability. The AI system was trained exclusively on data from a single clinical center. Although we subsequently validated its performance on two additional multi-site cohorts and an exploratory prospective set, these validation cohorts were of modest size. Therefore, this single-center training constraint, combined with the exploratory scale of the validation, means the system’s acceptable generalization ability to vastly different imaging software, operational protocols, and diverse patient populations is not yet fully proven. Second, regarding model performance, the recall of the INFLAM task was suboptimal when operating in a standalone mode. However, the human-AI integrated mode significantly restored recall to clinically robust levels, validating the necessity of the “human-in-the-loop” workflow. More fundamentally, a third limitation concerns the scope of our assessment. Our proposed system exclusively evaluates radiographic determinants strongly associated with esthetic risk (i.e., INFLAM, RESTOR, and DISTAN), which are critical pre-surgical proxies. However, direct clinical esthetic outcomes—such as the Pink Esthetic Score (PES), White Esthetic Score (WES), or patient-reported outcome measurements were not measured or prospectively validated in this study. We must acknowledge that the macro-level goal of predicting the final, patient-centered esthetic outcome is exceptionally complex. Holistic esthetic assessment inherently involves three-dimensional anatomical soft tissue profiles, dynamic facial harmony, and subjective patient perception, none of which can be fully captured by 2D radiographic screening alone. Therefore, future large-scale prospective cohorts must aim to correlate these AI-generated radiographic risk classifications directly with clinical esthetic endpoints to comprehensively validate the system’s prognostic impact. Therefore, future directions must prioritize critical pathways to address these limitations. First, a large-scale, definitive multi-center prospective validation study is essential to rigorously test and conclusively confirm the system’s robustness and generalization ability across real-world hardware heterogeneity. Second, the development of adaptive classification networks trained on larger, more diverse datasets will be necessary to propel this tool from a promising prototype to a universally reliable clinical solution. Third, future work must move from “risk assessment” toward “holistic assessment” by integrating multi-modal data. This includes developing adaptive networks trained on larger, more diverse datasets that incorporate clinical photographs, patient-reported outcome measures, and other patient-specific factors. The integration of the AI system into dental radiography-based esthetic outcome risk assessment, particularly through sophisticated systems like CNN, represents a promising clinical applicability in dental esthetic treatment planning. The system provides a more objective, reliable, and efficient means of assessing esthetic risk factors associated with dental implantation, enhancing the predictive capabilities of dental professionals. As AI technology continues to evolve, it holds the promise of further advancing dental care, making diagnostics more accessible, quicker, and potentially more accurate, thereby improving patient outcomes and streamlining clinical workflows. Methods Sample collection and selection The experimental design incorporates two sequential stages: system development followed by a clinical validation test comprising the exploratory prospective clinical evaluation and external validation. The entire study was conducted in accordance with the Declaration of Helsinki and approved by the research committee and ethics committee of the Hospital of Stomatology, Sun Yat-sen University. The data for system development comprised periapical radiographies collected at the Department of Implantology, Hospital of Stomatology, Sun Yat-sen University, from 2013 to 2023. These images focused on single-tooth gaps in the anterior regions of both the maxillary and mandibular arches. Ethical approval was granted by the Ethics Committee of the Hospital of Stomatology, Sun Yat-sen University (Approval No: KQEC-2020-29-06). Due to the retrospective nature of this part, the requirement for informed consent was waived by the Ethics Committee, pursuant to the Measures for Ethical Review of Life Science and Medical Research Involving Humans (2023) issued by the National Health Commission of China, and consistent with the principles of the Declaration of Helsinki. To strictly enforce patient confidentiality, a multi-layered de-identification pipeline was implemented. All radiographs and clinical records were first processed by an independent clinician not involved in the subsequent AI analysis, who replaced all direct personal identifiers (e.g., names, medical record numbers) with non-informative, random alphanumeric codes. The master linkage key connecting these codes to patient identities was stored on a physically isolated, password-protected server, accessible only to the principal investigator for necessary auditing. Furthermore, regarding the radiographic images, a dual-level masking protocol was applied: (1) all metadata tags were automatically stripped, and (2) manual visual inspection was performed to identify and digitally crop or mask to prevent any potential identity leakage. The research team was provided access exclusively to this fully anonymized dataset. To ensure strict statistical independence between samples and to eliminate potential intra-patient correlation bias, the validation datasets, both exploratory prospective and multi-site cohorts, were meticulously curated so that exactly one independent periapical radiograph was included per patient. The inclusion criteria were (1) adults aged 18 years and above, (2) radiographs taken by professionally trained technicians, (3) single-tooth gaps in anterior regions, and (4) complete clinical records. Exclusion criteria included (1) spaces with multiple missing teeth; (2) images with poor quality; (3) unclear images caused by factors, such as those caused by metal, motion, or insufficient soft tissue contrast; or (4) radiographs exhibiting significant geometric distortion. Image acquisition and quality control To minimize geometric distortion and ensure the highest fidelity of radiographic features, all periapical radiographs were acquired using the standardized long-cone paralleling technique. The use of rigid beam alignment devices was mandatory to ensure that the digital sensor remained strictly parallel to the long axis of the tooth while the central X-ray beam was directed perpendicularly to both. This geometric arrangement serves as the gold standard for minimizing elongation or foreshortening artifacts. The images were captured using three standardized intraoral radiography units (X-Mind™ DC, Acteon; or Soredex Minray) by a team of certified oral radiology technicians. Strict adherence to institutional exposure protocols was enforced to maintain consistent image density and contrast across different acquisition units, thereby reducing scanner-induced variability in the dataset. To ensure the dataset met stringent research standards beyond routine clinical requirements, we implemented a rigorous three-tiered quality control framework to exclude images with non-comparable angulation or artifacts. (1) Level 1 (Technical): A real-time assessment was performed by technicians immediately post-acquisition to reject and retake images exhibiting gross technical errors, such as cone-cutting, motion artifacts, or severe angulation deviations. (2) Level 2 (Clinical): Attending clinicians verified the diagnostic value of the images; radiographs deemed insufficient for clinical decision-making were excluded. (3) Level 3 (Research Curation): The research team conducted a final, retrospective review to identify and exclude images with subtle geometric distortions (e.g., significant cusp separation or distorted crown-to-root ratios) that might have passed clinical screening but were unsuitable for precise AI training. For the DISTAN task, which involves quantitative distance measurement, specific measures were taken to prevent measurement bias caused by variable sensor-to-source distances. We utilized the intrinsic pixel-to-millimeter calibration factor derived from the metadata of the digital sensor system. All physical distances, specifically the vertical distance from the contact point to the alveolar bone crest, were calculated algorithmically using this calibration data. This approach mathematically substantially mitigates magnification errors inherent to 2D radiography, ensuring that the binary classification relied on objective physical dimensions rather than subjective visual estimation. Image annotation and gold standard acquisition The key esthetic predictive factors for the esthetic zone on each periapical radiography were marked as: infection in adjacent teeth (INFLAM task), existing restorations in adjacent teeth (RESTOR task), and risk associated with the distance from the contact points of adjacent teeth to the alveolar crest (DISTAN task). The tasks were classified as binary: # for INFLAM task: Favorable outcome: absence of a low-density shadow larger than 1 mm at the implant or apical sites of adjacent teeth; Favorable outcome: absence of a low-density shadow larger than 1 mm at the implant or apical sites of adjacent teeth; Suboptimal outcome: presence of a low-density shadow larger than 1 mm at the implant or apical sites of adjacent teeth. Suboptimal outcome: presence of a low-density shadow larger than 1 mm at the implant or apical sites of adjacent teeth. # for RESTOR task: Favorable outcome: absence of restoration images in adjacent teeth or root canals; Favorable outcome: absence of restoration images in adjacent teeth or root canals; Suboptimal outcome: high-density restoration images in adjacent teeth or root canals. Suboptimal outcome: high-density restoration images in adjacent teeth or root canals. # for DISTAN task: Favorable outcome: distance <7 mm between the contact point and the alveolar ridge; Favorable outcome: distance <7 mm between the contact point and the alveolar ridge; Suboptimal outcome: distance ≥7 mm between the contact point and the alveolar ridge. Suboptimal outcome: distance ≥7 mm between the contact point and the alveolar ridge. Measurements were conducted using ImageJ Fiji (National Institutes of Health, Bethesda, MD, USA), and annotations were performed using LabelMe (version 5.1.1)37. Two experienced dentists performed annotations with over 5 years of clinical experience. Prior to final classification, measurement reproducibility was rigorously ensured through formal intra-rater and inter-rater reliability testing. A third, senior expert dentist with over 10 years of experience, then served as an arbitrator, reviewing all discordant annotations to establish the final gold standard. For the DISTAN task, both human annotators and the AI system initially calculated a precise, continuous physical distance in millimeters utilizing the intrinsic pixel-to-millimeter calibration factor. The clinical dichotomization threshold into favorable (<7 mm) and suboptimal (≥7 mm) outcomes was strictly applied only subsequent to this continuous measurement. AI system architecture and multi-task pipeline The system is designed to address the multi-task objective of analyzing dental periapical radiographies, ensuring precise infection detection, restoration classification, and distance measurement. It integrates three specialized modules for detection, classification, and segmentation to solve different sub-tasks, achieving seamless integration in this system. For the INFLAM & RESTOR task, the detection module identifies regions of interest from the raw input images, which are further cropped from the original image and fed into the classification module for final binary assessment. For the DISTAN task, a segmentation module was employed to delineate the edentulous region. The segmentation masks were further processed using the Douglas-Peucker algorithm to approximate an ideal quadrilateral, from which key points were extracted to estimate the distance value. These tasks are designed to work synergistically within a unified pipeline (Fig. 8). Schematic of the AI system design for evaluating radiographic surrogate markers of implant esthetic risk. The INFLAM and RESTOR tasks use EfficientDet to detect regions of interest, followed by DenseNet classification into favorable or suboptimal. The DISTAN task applies nnU-Net segmentation with post-processing to measure the contact point–alveolar crest distance. Each task outputs binary assessments integrated for the final outcome risk assessment. Training, validation, and testing of an AI system Our model selection rationale was guided by a combination of empirical knowledge and experimental results. The scarcity of annotated medical imaging data precluded the use of pre-training models. Consequently, we prioritized convolutional neural network (CNN) architectures renowned for their ability to converge effectively on smaller datasets. For the segmentation module, we adopted nnU-Net38 framework for its robust auto-configuring capability. We followed the default configuration with a convolutional U-Net with 8 layers of encoders and decoders to efficiently extract and fuse rich multi-scale features. For the detection module, we utilized the mmDetection framework to benchmark four state-of-the-art models via cross-validation, ultimately selecting EfficientDet39 for its superior balance of accuracy and efficiency. Similarly, for the classification stage, we conducted cross-validation on five different backbones, with DenseNet40 demonstrating the best performance for our specific tasks. The dataset comprises 318 periapical radiographs, randomly split into an 80% development set and a 20% held-out test set. All model training, hyperparameter tuning, and five-fold cross-validation were conducted exclusively within the development set. The test set remained fully isolated and was used only once for final evaluation. Prior to training, all images were resized to 512 × 512 pixels and normalized based on the training set’s mean and standard deviation. To improve model robustness and prevent overfitting, data augmentation techniques were applied on-the-fly to the training set, including random rotation, scaling, flipping, and brightness and contrast jitter. During the training process, the development set was randomly divided into five folds, and the five-fold cross-validation strategy was employed for model selection and hyperparameter optimization. In each training, four folds of data were used for training, and the remaining fold served as an internal validation set. Hyperparameters (including learning rate, batch size, weight decay, and number of training epochs) were optimized via grid search based on validation performance, with early stopping applied to mitigate overfitting. Based on the cross-validation results, the optimal hyperparameters and model configurations were determined. A final model for each module was then retrained using the entire development set under the selected settings. All modules (segmentation, detection, and classification) were trained and optimized independently to maximize task-specific performance and were subsequently integrated into a sequential inference pipeline. For each individual module, the SGD optimizer with batch size 8 and weight decay 0.0001 was used to train the model. The learning rate starts from 0.001 and changes with a cosine annealing schedule. The maximum number of epochs was set to 75 for the detection and classification modules, and 100 for the segmentation module, to align with their respective training convergence speeds. During model development, AUROC curves were generated on the validation folds to assess alternative operating points. The classification threshold was determined using Youden’s Index based on validation performance, and this threshold was fixed prior to evaluation on the test set. The held-out test set was then used to evaluate the generalization capabilities of the models on unseen data. Model performance was comprehensively assessed using metrics including accuracy, precision, recall, F1-score, AUROC and AUPRC. The entire pipeline was implemented in Python (version 3.9.22) and executed on a platform equipped with one NVIDIA GeForce RTX 3090 GPU, ensuring computational efficiency and robust performance. Comparison and integration between dentists and the system Six dentists participated in this study, including two expert dentists with over five years of experience, two senior dentists with less than five years of experience, and two junior dentists without clinical experience. The dentists assessed the presence of adjacent tooth infections and restorations through visual inspection, while distance measurements were conducted using ImageJ Fiji (National Institutes of Health, Bethesda, MD, USA). A hybrid human-machine integration approach was introduced to combine the dentists’ expertise with an AI system. In the human-machine integration experiment, a consensus rule was applied: if either the human practitioner or the AI system identified a case as positive (i.e., with infection, with restoration, >7 mm distance), it was considered positive. This “OR” integration rule was intentionally selected to maximize diagnostic sensitivity. This high-sensitivity approach prioritizes the minimization of false negatives (i.e., missed at-risk cases), which is a critical objective in a clinical screening and decision-support context. The accuracy, recall, specificity, precision and F1-score of dental practitioners, both with and without integration with the AI system, are presented. Exploratory prospective clinical evaluation test of the AI system Periapical radiographies for an exploratory prospective clinical evaluation test were collected at the Department of Implantology, Hospital of Stomatology, Sun Yat-sen University, from January 2024 to May 2025 (Approved No: KQEC-2025-058-01), and all participants provided written informed consent to participate. These images, which primarily focused on single-tooth gaps in the anterior regions of both the maxillary and mandibular arches, were selected for exploratory prospective AI analysis. The inclusion criteria and exclusion criteria were the same as those described in “Methods: Sample collection and selection”. During this step, 49 new periapical radiographs were collected for further evaluation. Annotation of the gold standard followed the same process described in “Methods: Image annotation and gold standard acquisition”. Generalization ability and multi-site validation of the AI system Periapical radiographies for the multi-site validation were collected from two additional, operationally independent clinics affiliated with the Hospital of Stomatology, Sun Yat-sen University: the Zhujiang New Town clinic and the Dongpu clinic, with data spanning from January 2017 to December 2023. These satellite clinics are geographically separate from the main hospital and utilize different imaging equipment, making them a robust test of the model’s generalization ability. Data collection from these sites was covered under the same institutional ethical approval (Approval No: KQEC-2020-29-06) as part of the same hospital system. Consistent with the development dataset, the requirement for informed consent was waived by the Ethics Committee for this retrospective data collection. The inclusion criteria and exclusion criteria were the same as those described in “Methods: Sample collection and selection”. During this validation step, 27 new periapical radiographs were collected in total. Annotation of the gold standard for these multi-site cases followed the same process as described in “Methods: Image annotation and gold standard acquisition”. Statistical analysis Inter-rater agreement between the two initial dentists before arbitration was assessed using Cohen’s Kappa. Statistical analysis was conducted using SPSS (version 26.0, IBM, USA), and continuous variables were summarized as means and standard deviations, while categorical variables were presented as frequencies and percentages. Key performance metrics (e.g., accuracy, precision, recall, etc.) are reported with their 95% confidence intervals. Comparisons of performance between the AI system and dentists, as well as between dentists with and without AI assistance, were performed using McNemar’s test. Statistical significance was generally defined as a two-sided p < 0.05. However, to control for Type I errors during multiple comparisons, a Bonferroni correction was applied to adjust this significance threshold. Results were visualized using bar charts, radar plots and confusion matrices to demonstrate efficacy and reliability. Acknowledgements This work was financially supported by National Natural Science Foundation of China (82402380), Young Science and Technology Talent Support Program of Guangdong Precision Medicine Application Association (YSTTGDPMAA202502), Guangzhou Science and Technology Program key projects (2023B03J1232), Clinical Research Project of Guangdong Provincial Clinical Research Center of Oral Diseases, Construction Project of the Degree and Postgraduate Education Reform of Sun Yat-Sen University, Postgraduate Innovation Program of Sun Yat-Sen University and High Education Teaching Research Project of Guangdong Province (25GZD001), Undergraduate Training Program for Innovation of Sun Yat-sen University (20250584). We would like to thank Deyi Liu for assistance with algorithm writing, as well as Yingying Liang and Shuanglin Nie for their help in data collection. Author contributions H.L. and Z.G. conceived and designed the study, contributed to data collection and analysis, performed the statistical analysis, and were responsible for drafting the manuscript. B.W. contributed to algorithm writing, algorithm testing and critically revised the manuscript. L.Q., X.M., G.C., P.Z., and S.C. contributed to data acquisition and analysis. M.S. contributed to conception, design, data acquisition and critically revised the manuscript. X.Z. and Z.C. contributed to conception, design and critically revised the manuscript. R.W. contributed to conception, design, algorithm guidance, and critically revised the manuscript. Z.C. contributed to conception, design, data acquisition, analysis, and interpretation, and critically revised the manuscript. All authors read and approved the final manuscript. Data availability The datasets generated and/or analyzed during the current study are not publicly available due to patient privacy and ethical restrictions, but are available from the corresponding author on reasonable request. Code availability The code for this study is open source and has been released at https://github.com/wenbc21/EsthRecon-PeriRad. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Hengyi Liu, Zhuohong Gong, Beichen Wen. Contributor Information Zhuofan Chen, Email: chzhuof@mail.sysu.edu.cn. Ruixuan Wang, Email: wangruix5@mail.sysu.edu.cn. Zetao Chen, Email: chenzet3@mail.sysu.edu.cn. References Associated Data Data Availability Statement The datasets generated and/or analyzed during the current study are not publicly available due to patient privacy and ethical restrictions, but are available from the corresponding author on reasonable request. The code for this study is open source and has been released at https://github.com/wenbc21/EsthRecon-PeriRad.