Deep learning models for determining the relationship between mandibular third molars and inferior alveolar canal: a systematic review and meta-analysis

Objective: This systematic review and meta-analysis aimed to investigate the current body of literature on the use of artificial intelligence (AI) for determining the relationship between the inferior alveolar canal (IAC) and mandibular third molars (M3s).

Methods: This systematic review was conducted following PRISMA-2020 guidelines and prospectively registered in Prospero (CRD42024521750). Comprehensive searches in PubMed, Scopus, Web of Science, Embase, and Cochrane were performed on February 25, 2025, using a predefined strategy targeting studies evaluating AI for preoperative detection of the relationship between the mandibular third molar and the inferior alveolar canal via panoramic radiographs or CBCT. Two independent reviewers screened articles, extracted data, and assessed quality using the QUADAS-2 tool. Diagnostic accuracy metrics were synthesized via univariate and bivariate random-effects meta-analysis.

Results: In this systematic review, the meta-analysis yielded a pooled diagnostic accuracy of 82.43% (95% CI: 80.56–84.23), with sensitivity at 83.78% (95% CI: 80.93–86.44) and specificity at 76.32% (95% CI: 69.81–82.26). Although patient selection bias was low, variability in the index test and reference standard was noted due to differences in imaging modalities. Subgroup analyses highlighted superior performance for VGG-19, ResNet-101, and ResNet50v2, while meta-regression revealed significant negative impacts of RetinaNet and SqueezeNet, amid high heterogeneity (I²=98.1%).

Conclusion: The meta-analysis revealed a pooled diagnostic accuracy of 82.43% with sensitivity at 83.78% and specificity at 76.32% for AI models predicting the IAC–M3 relationship. These robust pooled metrics highlight the potential of CNN-based techniques to enhance diagnostic precision in dental imaging.