Related Experiment Video
Updated: Aug 17, 2026
Reliability of Artificial Intelligence-Based Cone Beam Computed Tomography Integration with Digital Dental Images
Published on: February 23, 2024
Validation of artificial intelligence-assisted CBCT analysis for predicting inferior alveolar nerve proximity to
Huan Hu1, Jiahang Wu2, Hongji Pu3
1Department of Stomatology, Qujing Central Hospital of Yunnan Province, Qujing, Yunnan, 655000, China.
Objectives:
This study aimed to evaluate the diagnostic accuracy of an artificial intelligence (AI)-assisted cone-beam computed tomography (CBCT) analysis system for predicting the spatial proximity of the inferior alveolar nerve (IAN) to impacted mandibular third molars (M3M), using expert radiologist assessment as the reference standard.
Methods:
A retrospective diagnostic accuracy study was conducted on an internal institutional cohort of 312 patients (mean age 28.21 ± 6.62 years; January 2021-December 2024). A deep learning system employing a modified U-Net architecture automatically segmented the IAN canal and M3M on CBCT and classified the IAN-M3M spatial relationship into three categories: no contact (> 2 mm), proximity (0-2 mm), and contact/overlap. Diagnostic accuracy was evaluated on an independent hold-out test set of 486 M3 sites (283 patients). Two senior oral and maxillofacial radiologists provided the reference standard. Sensitivity, specificity, PPV, NPV, AUC, and Cohen's kappa were calculated; 95% CIs were derived by Wilson score method (proportions) and bootstrap resampling (AUC, kappa).
Results:
The AI system achieved an overall accuracy of 90.1% (438/486; 95% CI: 87.1-92.5%), weighted AUC of 0.925 (95% CI: 0.904-0.944), and Cohen's κ of 0.851 (95% CI: 0.821-0.912), indicating almost perfect agreement with the expert reference standard. Per-category sensitivity ranged from 88.2% to 91.2% and specificity from 92.8% to 97.4%. Bland-Altman analysis revealed a mean difference of 0.06 mm (95% LoA: -0.63 to 0.75 mm). Mean DSC was 0.90 ± 0.04 for IAN canal and 0.93 ± 0.03 for M3M segmentation. AI processing time was 4.75 ± 1.12 s versus 189.12 ± 41.99 s for expert assessment (39.79-fold reduction; P < 0.001). Subgroup analysis showed highest accuracy for mesioangular (93.1%) and horizontal (91.6%) impaction.
Conclusions:
The AI-assisted CBCT analysis system demonstrated high diagnostic accuracy and excellent agreement with expert radiological assessment in predicting IAN proximity to impacted mandibular third molars, while substantially reducing processing time. These results support the potential of AI-driven automated CBCT analysis as a preoperative decision-support tool; however, the reference standard was radiographic rather than intraoperative, and prospective multicenter validation incorporating surgical outcome data will be required before definitive clinical deployment recommendations can be made.
