Assessing the accuracy of artificial intelligence in mandibular canal segmentation compared to semi-automatic segmentation on cone-beam computed tomography images

Abstract

Purpose This study aims to assess the accuracy of artificial intelligence (AI) in mandibular canal (MC) segmentation on cone-beam computed tomography (CBCT) compared to semi-automatic segmentation. The impact of third molar status (absent, erupted, impacted) on AI performance was also evaluated. Material and methods A total of 150 CBCT scans (300 MCs) were retrospectively analysed. Semi-automatic MC segmentation was performed by experts using Romexis software, serving as the reference standard. AI-based segmentation was conducted using Diagnocat, an AI-driven cloud-based platform. Three-dimensional segmentation accuracy was assessed by comparing AI and semi-automatic segmentations through surface-to-surface distance metrics in Cloud Compare software. Statistical analyses included the intraclass correlation coefficient (ICC) for inter- and intra-rater reliability, Kruskal-Wallis tests for group comparisons, and Mann-Whitney <i>U</i> tests for post-hoc analyses. Results The median deviation between AI and semi-automatic MC segmentation was 0.29 mm (SD: 0.25-0.37 mm), with 88% of cases within the clinically acceptable limit (≤ 0.50 mm). Inter-rater reliability for semi-automatic segmentation was 84.5%, while intra-rater reliability reached 95.5%. AI segmentation demonstrated the highest accuracy in scans without third molars (median deviation: 0.27 mm), followed by erupted third molars (0.28 mm) and impacted third molars (0.32 mm). Conclusions AI demonstrated high accuracy in MC segmentation, closely matching expert-guided semi-automatic segmentation. However, segmentation errors were more frequent in cases with impacted third molars, probably due to anatomical complexity. Further optimisation of AI models using diverse training datasets and multi-centre validation is recommended to enhance reliability in complex cases.