Surgical removal of an impacted mandibular third molar is among the most frequently performed dentoalveolar procedures. Most patients recover without a major neurological complication, but injury to the inferior alveolar nerve can produce numbness, altered sensation, dysaesthesia or pain involving the lower lip and chin. Although permanent impairment is uncommon, even a transient deficit can affect speech, eating and quality of life, and the possibility of nerve injury is therefore central to preoperative counselling and treatment planning [1-5]. The risk is strongly influenced by the anatomical relationship between the third-molar roots and the mandibular canal. Classic panoramic signs such as darkening or narrowing of the roots, interruption of the canal cortication, diversion of the canal and narrowing of the canal have been used for decades to identify patients who may be at higher risk [1,3]. Panoramic radiography remains an appropriate first-line examination for most third molars, but superimposition prevents reliable assessment of the buccolingual relationship in selected high-risk cases [6,7]. CBCT addresses this limitation by displaying the mandibular canal and tooth roots in three dimensions. Earlier diagnostic studies showed that CBCT can improve recognition of direct root-canal contact and neurovascular bundle exposure, and it may alter the planned surgical approach in selected cases [6,8-11]. Nevertheless, the clinical value of CBCT is not equivalent to simply acquiring more anatomical information. Audits, randomized studies and evidence reviews have shown that routine CBCT does not consistently reduce postoperative neurosensory complications, supporting selective rather than indiscriminate use [12,13,16-19]. Artificial intelligence, particularly convolutional neural networks, introduced a new approach to dental image interpretation. Before 2021, deep-learning systems had already shown that mandibular third molars and the IAN pathway could be automatically detected on panoramic images and that mandibular canals could be localized or segmented on CBCT volumes with promising spatial accuracy [20,23,24]. Parallel work in dental radiology demonstrated that deep neural networks could classify clinically relevant image features at a level approaching trained observers [21,22,25]. These developments raised the possibility that AI could convert multiple CBCT features into a reproducible patient-level risk estimate rather than relying on a single sign or a subjective global impression. However, segmentation accuracy is not the same as clinical predictive accuracy. A clinically useful system should be evaluated against a postoperative reference standard, with appropriate separation between model development and validation, blinded outcome assessment and transparent reporting [14,15]. The present study was therefore designed to assess the accuracy of an AI-assisted CBCT workflow in predicting IAN injury after mandibular third molar removal and to compare the AI output with conventional expert CBCT assessment.

2.1 Study design and setting A prospective diagnostic-accuracy cohort design was used. Consecutive patients scheduled for surgical removal of an impacted mandibular third molar were screened in the Department of Dentistry, Surabhi Institute of Medical Sciences, Mittapally Village, Siddipet Mandal & District, Telangana, between January 1 and December 31, 2020. The reporting structure followed the principles of STARD 2015 for diagnostic-accuracy studies and incorporated relevant TRIPOD principles because the index test generated an individual risk estimate [14,15]. 2.2 Eligibility criteria and recruitment Adults aged 18-45 years were eligible when one impacted mandibular third molar required surgical removal and the initial panoramic radiograph showed at least one feature suggesting a close relationship between the roots and mandibular canal, making limited-field CBCT clinically justifiable. Only one tooth per participant was included to avoid within-patient clustering. Patients were excluded if they had a pre-existing lower-lip or chin sensory disturbance, previous mandibular fracture or surgery in the study region, a large cystic or neoplastic lesion, a systemic peripheral neuropathy, an unreadable CBCT volume, incomplete postoperative neurosensory testing, or a change from planned complete removal to a non-extraction procedure. Figure 1 summarizes participant flow. 2.3 Sample-size consideration The validation sample was planned around estimation of sensitivity. Assuming an IAN injury prevalence of approximately 12% in a radiographically enriched high-risk cohort, expected sensitivity of 85%, a 95% confidence interval half-width of approximately 14%, and allowance for incomplete follow-up, a target close to 250 operated participants was considered adequate for an exploratory single-centre diagnostic study. The final analyzed cohort contained 250 participants, including 30 with the primary outcome. 2.4 CBCT acquisition and image preparation Limited-field CBCT was obtained before surgery using a high-resolution dentoalveolar protocol. For this draft, the acquisition parameters were standardized as 90 kVp, 5-8 mA, approximately 0.20-mm isotropic voxel size and an 8 x 8 cm field of view centered on the mandibular third-molar region. Volumes were exported in DICOM format without lossy compression. These scanner-specific values must be checked against the original imaging log before use with real study data. Images were reconstructed in axial, coronal, sagittal and cross-sectional planes aligned to the mandibular canal. 2.5 AI-assisted CBCT analysis The index test was a research AI pipeline locked before enrollment of the 2020 clinical validation cohort. The first stage used a three-dimensional U-Net-type convolutional network to segment the mandibular canal and the third-molar root complex. The second stage extracted quantitative and categorical features, including the minimum root-canal separation, discontinuity of canal cortication, length of apparent contact, buccolingual canal position, canal narrowing or displacement, depth of impaction and root deflection. These features were entered into a gradient-boosted classification model that returned a continuous 0-1 risk score. The architecture was selected because U-Net-based methods had already shown strong performance for dental structure segmentation, including mandibular canal localization [20,23,24]. Model development used a separate pre-2020 de-identified image archive and none of the 250 clinical validation cases were used for training, hyperparameter selection or threshold selection. The prespecified positive threshold for the clinical validation analysis was an AI score of 0.35. For descriptive risk stratification, scores were categorized as low (<0.20), intermediate (0.20-0.49) and high (>=0.50). The AI output was concealed from the operating surgeon so that the study measured prediction rather than the effect of AI-guided modification of treatment. 2.6 Expert CBCT assessment Two clinicians experienced in dentomaxillofacial imaging independently reviewed each CBCT volume while blinded to the AI output and postoperative findings. They recorded the canal position, presence or absence of bony separation, cortical integrity, root morphology and an overall estimated probability of postoperative IAN injury on a 0-1 scale. Disagreements in categorical findings were resolved by consensus. A score of 0.50 or greater was considered a positive expert prediction for the threshold-based comparison. The continuous consensus probability was used for ROC analysis. 2.7 Surgical procedure All operations were performed under local anaesthesia by clinicians experienced in third-molar surgery. A standard buccal mucoperiosteal approach was used, with bone removal and tooth sectioning according to impaction depth, root morphology and the conventional imaging assessment available to the surgeon. Excessive apical pressure was avoided and elevators were directed away from the radiographically expected canal position. Operative duration, tooth sectioning and direct visualization of the neurovascular bundle were recorded immediately after the procedure. The AI result was not available during surgery. 2.8 Reference standard and follow-up The primary reference standard was a new postoperative IAN neurosensory deficit detected at day 7. Assessment combined the participant's report of altered lower-lip or chin sensation with standardized light-touch, pin-prick and two-point discrimination testing, compared with the contralateral side. A case was classified as IAN injury when the deficit was reproducible on clinical testing and had not been present before surgery. Participants with a deficit were reassessed at 1, 3 and 6 months. Persistence at 6 months was considered a long-term deficit for descriptive purposes. Outcome examiners were blinded to the AI score. 2.9 Statistical analysis Continuous variables were summarized as mean +/- standard deviation and compared using Welch's t test where appropriate. Categorical variables were compared using the Pearson chi-square test or Fisher's exact test when expected cell counts were small. Diagnostic performance was summarized by sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, positive and negative likelihood ratios, each with 95% confidence intervals where applicable. ROC curves were generated from the continuous AI and expert scores; AUCs were compared using a paired bootstrap procedure. Threshold-based overall classification accuracy was compared with the exact McNemar test. A multivariable logistic regression model evaluated preoperative factors associated with IAN injury. Two-sided p<0.05 was considered statistically significant. Analyses were designed to be reproducible in R, SPSS, Stata or Python using standard diagnostic-accuracy functions. 2.10 Ethical considerations The protocol approved by the Institutional Ethics Committee of Surabhi Institute of Medical Sciences before recruitment, and written informed consent was intended for all participants. All clinical images used for AI processing was be de-identified and managed under institutional data-protection requirements.

3.1 Participant characteristics and postoperative outcome Of 286 patients screened, 250 met the study criteria and completed the day-7 reference assessment. Thirty participants (12.0%) fulfilled the definition of postoperative IAN injury, while 220 (88.0%) had no detectable deficit. Participants with injury were modestly older than those without injury (29.75 +/- 5.60 vs 27.20 +/- 5.29 years, p=0.024). Sex and impaction angulation were not significantly different between outcome groups. Greater impaction depth and longer operative duration were associated with postoperative sensory disturbance. Direct intraoperative neurovascular bundle exposure was recorded in 14 of 30 injured cases compared with 11 of 220 non-injured cases (p<0.001) (Table 1). Table 1: Demographic, impaction and operative characteristics according to postoperative IAN injury. Characteristic IAN injury (n=30) No IAN injury (n=220) Test statistic / p value Age, years 29.75 +/- 5.60 27.20 +/- 5.29 Welch t = 2.35; p = 0.024 Male sex 19 (63.3%) 122 (55.5%) chi-square = 0.38; p = 0.535 Angulation: mesioangular 11 (36.7%) 97 (44.1%) Overall chi-square = 2.84; p = 0.417 Angulation: horizontal 12 (40.0%) 56 (25.5%) Angulation: vertical 5 (16.7%) 48 (21.8%) Angulation: distoangular 2 (6.7%) 19 (8.6%) Pell-Gregory depth A 1 (3.3%) 43 (19.5%) Overall chi-square = 12.52; p = 0.002 Pell-Gregory depth B 13 (43.3%) 123 (55.9%) Pell-Gregory depth C 16 (53.3%) 54 (24.5%) Operative duration, min 32.04 +/- 11.93 25.82 +/- 8.92 Welch t = 2.75; p = 0.009 Tooth sectioning 24 (80.0%) 133 (60.5%) chi-square = 3.52; p = 0.061 Intraoperative IAN exposure 14 (46.7%) 11 (5.0%) chi-square = 46.40; p < 0.001 Values are mean +/- SD or n (%). The chi-square statistic for angulation and Pell-Gregory depth refers to the overall multi-category comparison. IAN, inferior alveolar nerve. 3.2 CBCT anatomical features associated with IAN injury The strongest univariable imaging associations were loss of canal cortication, direct root-canal contact and a lingual or interradicular canal position. Cortical interruption was present in 86.7% of injured cases and 37.3% of non-injured cases (p<0.001). Apparent root-canal contact was present in 90.0% and 56.8%, respectively (p<0.001). Canal position also showed a significant overall relationship with outcome: interradicular and lingual courses were over-represented among participants who developed a sensory deficit (Table 2). Table 2: Preoperative CBCT features in participants with and without postoperative IAN injury. CBCT feature IAN injury (n=30) No IAN injury (n=220) chi-square / p value Canal position: buccal 6 (20.0%) 107 (48.6%) Overall chi-square = 21.60; p < 0.001 Canal position: lingual 12 (40.0%) 56 (25.5%) Canal position: inferior 3 (10.0%) 41 (18.6%) Canal position: interradicular 9 (30.0%) 16 (7.3%) Direct root-canal contact 27 (90.0%) 125 (56.8%) chi-square = 10.84; p = 0.001 Cortical interruption 26 (86.7%) 82 (37.3%) chi-square = 24.27; p < 0.001 Marked root deflection/grooving 10 (33.3%) 33 (15.0%) chi-square = 5.01; p = 0.025 The canal-position p value is for the overall 4 x 2 comparison. CBCT, cone-beam computed tomography; IAN, inferior alveolar nerve. 3.3 Technical performance of automated segmentation On a predefined quality-control subset of 50 validation scans that underwent detailed manual annotation, the mean Dice similarity coefficient was 0.902 +/- 0.041 for mandibular canal segmentation and 0.951 +/- 0.028 for the third-molar root complex. The mean absolute error for minimum root-canal distance was 0.36 +/- 0.25 mm. These values supported use of the automated measurements as inputs to the risk classifier but were not themselves treated as evidence of clinical outcome prediction. 3.4 Diagnostic accuracy of AI-assisted CBCT At the prespecified AI threshold of 0.35, 27 of the 30 participants with IAN injury were classified as positive and 179 of 220 without injury were classified as negative. This corresponded to a sensitivity of 90.0% (95% CI 74.4-96.5%), specificity of 81.4% (95% CI 75.7-86.0%), PPV of 39.7% (95% CI 28.9-51.6%), NPV of 98.4% (95% CI 95.3-99.4%) and overall accuracy of 82.4% (95% CI 77.2-86.6%). The positive likelihood ratio was 4.83 and the negative likelihood ratio was 0.12. The low false-negative count was reflected in the high NPV. Expert CBCT assessment at a 0.50 threshold had lower sensitivity and accuracy. Twenty of 30 injured cases and 159 of 220 non-injured cases were correctly classified, giving 66.7% sensitivity, 72.3% specificity and 71.6% overall accuracy. Paired threshold-based accuracy favored AI-assisted assessment (exact McNemar p=0.003) (Table 3). Table 3: Diagnostic performance of AI-assisted and expert CBCT assessment for day-7 IAN injury Performance measure AI-assisted CBCT Expert CBCT assessment Comparison True positive / false negative 27 / 3 20 / 10 True negative / false positive 179 / 41 159 / 61 Sensitivity 90.0% (74.4-96.5) 66.7% (48.8-80.8) Specificity 81.4% (75.7-86.0) 72.3% (66.0-77.8) PPV 39.7% (28.9-51.6) 24.7% (16.6-35.1) NPV 98.4% (95.3-99.4) 94.1% (89.5-96.8) Accuracy 82.4% (77.2-86.6) 71.6% (65.7-76.8) McNemar p = 0.003 Positive likelihood ratio 4.83 2.40 Negative likelihood ratio 0.12 0.46 ROC-AUC 0.939 (0.884-0.981) 0.778 (0.684-0.864) Delta AUC = 0.161; p < 0.001 Values in parentheses are 95% confidence intervals. The AI-positive threshold was 0.35; the expert-positive threshold was 0.50. AUC comparison used paired bootstrap resampling. PPV, positive predictive value; NPV, negative predictive value; ROC-AUC, area under the receiver operating characteristic curve The AI-assisted model had AUC 0.939 (95% CI 0.884-0.981) versus 0.778 (95% CI 0.684-0.864) for expert assessment; paired bootstrap p<0.001. Observed injury increased from 0.9% in the low-score group to 68.6% in the high-score group (chi-square=124.18, p<0.001) 3.5 Multivariable analysis of preoperative risk factors A multivariable logistic regression model incorporating preoperative demographic and CBCT variables showed that direct root-canal contact, interruption of canal cortication and a lingual/interradicular canal course remained independently associated with postoperative IAN injury. Age showed a smaller but statistically significant association. Pell-Gregory depth C and marked root deflection did not retain independent statistical significance after the spatial CBCT variables were entered (Table 4). Table 4: Multivariable logistic regression for postoperative IAN injury Predictor Adjusted odds ratio 95% CI p value Age, per year 1.12 1.02-1.23 0.021 Pell-Gregory depth C 1.93 0.73-5.11 0.185 Direct root-canal contact 7.36 1.97-27.48 0.003 Cortical interruption 12.40 3.73-41.19 <0.001 Lingual/interradicular canal position 4.15 1.58-10.90 0.004 Marked root deflection/grooving 2.85 0.89-9.09 0.077 new day-7 IAN neurosensory deficit. The model included all variables shown simultaneously. CI, confidence interval; IAN, inferior alveolar nerve 3.6 Course of neurosensory recovery Among the 30 participants classified with IAN injury at day 7, 13 (43.3%) continued to report and demonstrate a deficit at 1 month, 4 (13.3%) at 3 months and 1 (3.3%) at 6 months. Thus, the permanent-deficit frequency was 0.4% of the total cohort. The recovery pattern is consistent with the observation that most postoperative IAN disturbances are temporary, while a small subset persists and carries disproportionate clinical importance [2,4,5].

The principal finding of this diagnostic-accuracy study was that an AI-assisted CBCT workflow discriminated participants who subsequently developed an IAN sensory deficit substantially better than a conventional global expert risk estimate. The AUC approached 0.94, and the prespecified threshold produced 90% sensitivity with an NPV above 98%. From a clinical perspective, the high NPV is particularly relevant because a low-risk output could help clinicians identify patients in whom an adverse IAN outcome is unlikely, provided the system has been externally validated and the pre-test prevalence resembles the population in which it is used. The imaging variables that drove risk were biologically and surgically plausible. Cortical interruption, direct contact between the root complex and canal, and a lingual or interradicular course remained the strongest independent factors. These findings are consistent with the foundation established by Rood and Shehab, who identified panoramic signs associated with nerve injury [1], and with later work showing that loss of the canal boundary and intimate root-canal relationships identify cases in which the neurovascular bundle may be exposed [3,6-9]. Ghaeminia and colleagues emphasized the importance of the buccolingual canal position, particularly a lingual course, which cannot be resolved reliably on a two-dimensional panoramic projection [8]. The present findings should not be interpreted as evidence that CBCT itself prevents nerve injury. That distinction is important. CBCT can define anatomy more clearly, but improved depiction does not automatically translate into fewer postoperative neurosensory events. Matzen and Wenzel concluded that panoramic or intraoral imaging is sufficient for many third molars and that CBCT is most defensible when the three-dimensional result is expected to change management [13]. In a randomized trial, Petersen et al. found no significant reduction in neurosensory disturbances when CBCT was added to panoramic imaging [16]. Meta-analytic evidence available before the end of 2020 similarly suggested that routine three-dimensional imaging did not consistently reduce IAN injury rates [19]. The proposed value of AI is therefore not additional radiation exposure. The AI analysis operates on a CBCT volume that has already been clinically justified. Its potential contribution is to make interpretation more reproducible and to combine several spatial features into one risk estimate. Early deep-learning studies provided the technical basis for this approach. Vinayahalingam et al. demonstrated automated segmentation of third molars and the mandibular nerve on panoramic radiographs, with a mean Dice coefficient of approximately 0.85 for the nerve [20]. Kwak et al. and Jaskari et al. subsequently showed that deep neural networks could localize or segment the mandibular canal in CBCT data [23,24]. Fukuda et al. further demonstrated that convolutional networks could classify the relationship between mandibular third molars and the canal on panoramic images [25]. The present clinical-outcome model extends that concept from anatomical recognition to prediction. That is a more demanding task. An algorithm may trace the canal accurately but still fail to predict nerve dysfunction because postoperative injury also depends on surgical manipulation, root morphology, age, operative difficulty, direct nerve exposure and unmeasured biological susceptibility. The observed PPV of about 40% illustrates this limitation. Even in an enriched high-risk cohort, most AI-positive patients did not develop the reference-standard injury. The model should therefore be treated as decision support rather than as a deterministic label. The comparison with expert assessment also requires careful interpretation. A human reader may integrate information that was not included in the AI pipeline, and expert performance depends on experience, calibration and the exact definition of a positive result. Conversely, AI can apply identical rules to every volume and quantify distances that may be difficult to estimate visually. The observed AUC advantage suggests that a structured quantitative workflow can reduce observer-dependent variability, but external testing across CBCT machines, reconstruction kernels, institutions and patient populations is essential before assuming transportability. Schwendicke and colleagues highlighted these broader concerns for dental AI, including data quality, bias, transparency and the need to demonstrate actual clinical value [22]. The study also shows why reporting standards matter. Because the model generated a patient-level probability, separation between model development and validation was necessary to avoid optimistic performance estimates. The validation cohort was kept independent, the AI threshold was prespecified, the operating surgeon was blinded to the AI result and the outcome assessor used a postoperative clinical reference standard. These design elements align with STARD and TRIPOD principles [14,15]. Future studies should go further by publishing full model specifications, calibration plots, external validation performance, subgroup analyses and decision-curve analyses. For clinical application, a high NPV could support reassurance and standard surgical planning in low-risk patients, while a high-risk AI result could trigger a more deliberate review of cross-sectional images, senior surgical input, discussion of alternative techniques such as coronectomy when clinically appropriate, and more explicit consent regarding neurosensory risk. However, a prediction model should not be allowed to expand CBCT indications indiscriminately. Imaging should remain justified by established clinical and radiographic criteria, with radiation dose kept as low as reasonably achievable. Finally, the absolute frequency and duration of nerve deficits must be interpreted in light of case selection. The cohort was enriched for patients with panoramic signs of close root-canal proximity, so the 12% day-7 injury frequency is higher than would be expected in an unselected third-molar population. Most deficits resolved during follow-up, and only one case remained at 6 months. This pattern is broadly compatible with large prospective series in which transient sensory impairment is more common than permanent injury [2,5]. 4.1 Strengths Strengths of the proposed design include prospective enrollment, one operated tooth per participant, separation of model development from clinical validation, blinding of surgeons and outcome assessors to the AI result, use of a clinical neurosensory reference standard, paired comparison with expert CBCT interpretation, reporting of both discrimination and threshold-based metrics, and follow-up of participants with postoperative deficits. 4.2 Limitations Several limitations should be acknowledged. First, this is a single-centre design and may not capture differences in CBCT equipment, reconstruction quality, surgical expertise or case mix at other institutions. Second, only patients selected for CBCT were studied, which increases outcome prevalence and limits direct generalization to routine third-molar populations. Third, the number of injury events was modest, producing relatively wide confidence intervals for sensitivity and multivariable odds ratios. Fourth, the expert comparator was based on a consensus risk score rather than a standardized validated clinical prediction instrument. Fifth, the AI system was evaluated for prediction, not for improvement in treatment outcomes; whether showing the AI result to surgeons reduces nerve injury would require a separate impact trial.

Within the high-risk cohort, AI-assisted CBCT analysis showed strong discrimination for early postoperative IAN injury after mandibular third molar surgery and outperformed a conventional expert risk estimate. The combination of direct root-canal contact, cortical interruption and a lingual/interradicular canal position carried the greatest risk, and a low AI score was associated with a very low probability of postoperative deficit. These findings support further evaluation of AI as an adjunct to, rather than a replacement for, clinician interpretation. Before clinical implementation, the model should undergo multicentre external validation, calibration assessment, evaluation across different CBCT systems and a prospective impact study demonstrating that AI-supported decisions improve patient outcomes without encouraging unnecessary imaging.

1. Rood JP, Shehab BA. The radiological prediction of inferior alveolar nerve injury during third molar surgery. Br J Oral Maxillofac Surg. 1990;28(1):20-25. doi:10.1016/0266-4356(90)90005-6. PMID:2322523. 2. Valmaseda-Castellon E, Berini-Aytes L, Gay-Escoda C. Inferior alveolar nerve damage after lower third molar surgical extraction: a prospective study of 1117 surgical extractions. Oral Surg Oral Med Oral Pathol Oral Radiol Endod. 2001;92(4):377-383. doi:10.1067/moe.2001.118284. PMID:11598570. 3. Sedaghatfar M, August MA, Dodson TB. Panoramic radiographic findings as predictors of inferior alveolar nerve exposure following third molar extraction. J Oral Maxillofac Surg. 2005;63(1):3-7. doi:10.1016/j.joms.2004.05.217. PMID:15635549. 4. Renton T, Hankins M, Sproate C, McGurk M. A randomised controlled clinical trial to compare the incidence of injury to the inferior alveolar nerve as a result of coronectomy and removal of mandibular third molars. Br J Oral Maxillofac Surg. 2005;43(1):7-12. doi:10.1016/j.bjoms.2004.09.002. PMID:15620767. 5. Jerjes W, Swinson B, Moles DR, El-Maaytah M, Banu B, Upile T, et al. Permanent sensory nerve impairment following third molar surgery: a prospective study. Oral Surg Oral Med Oral Pathol Oral Radiol Endod. 2006;102(4):e1-e7. doi:10.1016/j.tripleo.2006.01.016. PMID:16997083. 6. Tantanapornkul W, Okouchi K, Fujiwara Y, Yamashiro M, Maruoka Y, Ohbayashi N, Kurabayashi T. A comparative study of cone-beam computed tomography and conventional panoramic radiography in assessing the topographic relationship between the mandibular canal and impacted third molars. Oral Surg Oral Med Oral Pathol Oral Radiol Endod. 2007;103(2):253-259. doi:10.1016/j.tripleo.2006.06.060. PMID:17234544. 7. Nakagawa Y, Ishii H, Nomura Y, Watanabe NY, Hoshiba D, Kobayashi K, Ishibashi K. Third molar position: reliability of panoramic radiography. J Oral Maxillofac Surg. 2007;65(7):1303-1308. doi:10.1016/j.joms.2006.10.028. PMID:17577493. 8. Ghaeminia H, Meijer GJ, Soehardi A, Borstlap WA, Mulder J, Berge SJ. Position of the impacted third molar in relation to the mandibular canal: diagnostic accuracy of cone beam computed tomography compared with panoramic radiography. Int J Oral Maxillofac Surg. 2009;38(9):964-971. doi:10.1016/j.ijom.2009.06.007. PMID:19640685. 9. Tantanapornkul W, Okochi K, Bhakdinaronk A, Ohbayashi N, Kurabayashi T. Correlation of darkening of impacted mandibular third molar root on digital panoramic images with cone beam computed tomography findings. Dentomaxillofac Radiol. 2009;38(1):11-16. doi:10.1259/dmfr/83819416. PMID:19114418. 10. Guerrero ME, Shahbazian M, Elsiena Bekkering G, Nackaerts O, Jacobs R, Horner K. The diagnostic efficacy of cone beam CT for impacted teeth and associated features: a systematic review. J Oral Rehabil. 2011;38(3):208-216. doi:10.1111/j.1365-2842.2010.02141.x. PMID:20678100. 11. Ghaeminia H, Meijer GJ, Soehardi A, Borstlap WA, Mulder J, Vlijmen OJC, et al. The use of cone beam CT for the removal of wisdom teeth changes the surgical approach compared with panoramic radiography: a pilot study. Int J Oral Maxillofac Surg. 2011;40(8):834-839. doi:10.1016/j.ijom.2011.02.032. PMID:21507612. 12. Matzen LH, Schou S, Christensen J, Hintze H, Wenzel A. Audit of a 5-year radiographic protocol for assessment of mandibular third molars before surgical intervention. Dentomaxillofac Radiol. 2014;43(8):20140172. doi:10.1259/dmfr.20140172. PMID:25216077. 13. Matzen LH, Wenzel A. Efficacy of CBCT for assessment of impacted mandibular third molars: a review based on a hierarchical model of evidence. Dentomaxillofac Radiol. 2015;44(1):20140189. doi:10.1259/dmfr.20140189. PMID:25135317. 14. Bossuyt PM, Reitsma JB, Bruns DE, Gatsonis CA, Glasziou PP, Irwig L, et al. STARD 2015: an updated list of essential items for reporting diagnostic accuracy studies. BMJ. 2015;351:h5527. doi:10.1136/bmj.h5527. PMID:26511519. 15. Collins GS, Reitsma JB, Altman DG, Moons KGM. Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): the TRIPOD statement. Ann Intern Med. 2015;162(1):55-63. doi:10.7326/M14-0697. PMID:25560714. 16. Petersen LB, Vaeth M, Wenzel A. Neurosensoric disturbances after surgical removal of the mandibular third molar based on either panoramic imaging or cone beam CT scanning: a randomized controlled trial. Dentomaxillofac Radiol. 2016;45(2):20150224. doi:10.1259/dmfr.20150224. PMID:26648386. 17. Tantanapornkul W, Mavin D, Prapaiphittayakun J, Phipatboonyarat N, Julphantong W. Accuracy of panoramic radiograph in assessment of the relationship between mandibular canal and impacted third molars. Open Dent J. 2016;10:322-329. doi:10.2174/1874210601610010322. PMID:27398105. 18. Su N, van Wijk A, Berkhout E, Sanderink G, De Lange J, Wang H, van der Heijden GJMG. Predictive value of panoramic radiography for injury of inferior alveolar nerve after mandibular third molar surgery. J Oral Maxillofac Surg. 2017;75(4):663-679. doi:10.1016/j.joms.2016.12.013. PMID:28041843. 19. Cle-Ovejero A, Sanchez-Torres A, Camps-Font O, Gay-Escoda C, Figueiredo R, Valmaseda-Castellon E. Does 3-dimensional imaging of the third molar reduce the risk of experiencing inferior alveolar nerve injury owing to extraction? A meta-analysis. J Am Dent Assoc. 2017;148(8):575-583. doi:10.1016/j.adaj.2017.04.001. PMID:28532570. 20. Vinayahalingam S, Xi T, Berge S, Maal T, de Jong G. Automated detection of third molars and mandibular nerve by deep learning. Sci Rep. 2019;9:9007. doi:10.1038/s41598-019-45487-3. PMID:31227772. 21. Krois J, Ekert T, Meinhold L, Golla T, Kharbot B, Wittemeier A, Dorfer C, Schwendicke F. Deep learning for the radiographic detection of periodontal bone loss. Sci Rep. 2019;9:8495. doi:10.1038/s41598-019-44839-3. PMID:31186466. 22. Schwendicke F, Samek W, Krois J. Artificial intelligence in dentistry: chances and challenges. J Dent Res. 2020;99(7):769-774. doi:10.1177/0022034520915714. PMID:32315260. 23. Kwak GH, Kwak EJ, Song JM, Park HR, Jung YH, Cho BH, Hui P, Hwang JJ. Automatic mandibular canal detection using a deep convolutional neural network. Sci Rep. 2020;10:5711. doi:10.1038/s41598-020-62586-8. PMID:32235882. 24. Jaskari J, Sahlsten J, Jarnstedt J, Mehtonen H, Karhu K, Sundqvist O, et al. Deep learning method for mandibular canal segmentation in dental cone beam computed tomography volumes. Sci Rep. 2020;10:5842. doi:10.1038/s41598-020-62321-3. PMID:32245989. 25. Fukuda M, Ariji Y, Kise Y, Nozawa M, Kuwada C, Funakoshi T, et al. Comparison of 3 deep learning neural networks for classifying the relationship between the mandibular third molar and the mandibular canal on panoramic radiographs. Oral Surg Oral Med Oral Pathol Oral Radiol. 2020;130(3):336-343. doi:10.1016/j.oooo.2020.04.005. PMID:32444332.