U-Net-based deep learning architecture for automated CBCT segmentation of the mandibular canal in dental implant treatment planning: A systematic review and meta-analysis.

STATEMENT OF PROBLEM: The manual segmentation of anatomic structures in cone beam computed tomography (CBCT) scans such as the mandibular canal is time-consuming, operator-dependent, and prone to variability, limiting efficiency and consistency in implant planning. Whether automated CBCT segmentation is more accurate remains unclear.

PURPOSE: The purpose of this systematic review and meta-analysis was to synthesize evidence on deep learning, primarily U-Net architectures, for automatic segmentation of the mandibular canal in CBCT imaging. The segmentation accuracy and clinical relevance for dental implant planning was evaluated.

MATERIAL AND

METHODS

A systematic review and meta-analysis were conducted by following the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) 2020 guidelines using the PubMed, Scopus, and the Cochrane Library databases. Eligible studies applied U-Net-based segmentation to CBCT scans for dental implants. Data on segmentation accuracy, Dice Similarity Coefficient [DSC], 95% Hausdorff Distance [HD], and Intersection over Union [IoU]), model architecture, validation strategies, and clinical applicability were extracted. Study quality was appraised using the Checklist for Artificial Intelligence in Medical Imaging (CLAIM) and QUADAS-2 tools. A random-effects meta-analysis using logit-transformed means and standard errors were performed for DSC and IoU, with results back-transformed to the proportion scale, while 95% HD was analyzed on its raw scale. A sensitivity analysis of DSC was conducted on overlapping studies reporting both DSC and IoU to verify metric consistency, and publication bias was examined using the Egger regression test and the trim-and-fill method.

RESULTS

Eight studies were quantitatively synthesized. The pooled DSC was 0.84 (95% CI: 0.68-0.92), IoU was 0.81 (95% CI: 0.50-0.95), and 95% HD was 0.89 mm (95% CI: -0.13 to 1.92 mm), all demonstrating high heterogeneity (I²>99%). Sensitivity analysis of DSC on overlapping studies confirmed internal consistency between DSC and IoU (pooled DSC=0.89; IoU=0.81). The Egger tests indicated no significant publication bias (P=.932 for DSC; P=.898 for IoU; P=.626 for 95% HD). Advanced architectures such as attention- and residual-based U-Nets showed superior accuracy, though external validation and explainability analyses were rarely reported.

CONCLUSIONS

U-Net-based deep learning models show a strong potential for automated CBCT segmentation of the mandibular canal, offering improved efficiency and accuracy. However, widespread clinical adoption requires standardized reporting, external validation, and explainability features to ensure trust and generalizability.