AI agreed with radiologists on nerve proximity in 90.1 percent of CBCT sites
A deep learning system that outlines the inferior alveolar nerve canal and the impacted third molar itself matched the relationship in 90.1 percent of 486 sites, and processed a case in 4.75 seconds against 189 seconds for an expert.
Agreement with radiologists in 90.1 percent of assessed sites
A diagnostic accuracy study was published in BMC Oral Health by H. Hu, J. Wu and H. Pu. They examined how reliably a deep learning system predicts the spatial relationship between the inferior alveolar nerve and an impacted lower third molar on CBCT scans.
The study is retrospective and was carried out on an internal cohort of 312 patients (mean age 28.21 ± 6.62 years) scanned between January 2021 and December 2024. The architecture is a modified U-Net; the system automatically segments the canal and the third molar and classifies the relationship into three categories, no contact (more than 2 mm), proximity (0 to 2 mm), and contact or overlap. Testing was carried out on a separate set of 486 sites in 283 patients, with two experienced oral and maxillofacial radiologists as the reference standard.
Overall accuracy was 90.1 percent (438 of 486; 95% CI 87.1 to 92.5), the weighted AUC 0.925 and Cohen's kappa 0.851. Sensitivity by category was between 88.2 and 91.2 percent, specificity between 92.8 and 97.4 percent. Segmentation was precise (DSC 0.90 ± 0.04 for the canal and 0.93 ± 0.03 for the molar), with a mean deviation of 0.06 mm.
The largest difference was in time: 4.75 ± 1.12 seconds per case against 189.12 ± 41.99 seconds for an expert. Accuracy was highest in mesioangular (93.1 percent) and horizontal (91.6 percent) impaction.
Limitations
The reference standard was radiological rather than intraoperative: the system was compared with how a radiologist reads the scan, not with what the surgeon finds. The cohort comes from a single institution, and the authors themselves call for prospective multicentre validation with surgical outcome data before recommending clinical use.
For the practice
- A tool like this helps with triage, it does not replace the report: where contact or overlap is found, the surgeon still decides, and the patient's consent must mention the risk to the nerve regardless of what the algorithm wrote.
- If you order CBCT with automatic segmentation, ask on which population the model was validated and whether the reference standard was radiological or surgical.
- The time saving is real, but it pays off only when it is clear who checks and signs the report.