Accurate assessment of the relationship between mandibular third molars and the mandibular canal is essential to reduce the risk of inferior alveolar nerve injury. Artificial intelligence (AI), including deep learning and convolutional neural network-based models, has been increasingly investigated using panoramic radiography and cone-beam computed tomography (CBCT) for automated diagnostic support in third molar assessment. A systematic review (SR) was conducted following the PECOS framework to answer the research question: “ Can artificial intelligence determine the relationship between the mandibular canal and mandibular third molars ?” Searches were performed in the PubMed, Scopus, CAPES Periodicals, LILACS, and Cochrane Library databases up to February 2026. Studies evaluating AI models for identifying or classifying the relationship between the mandibular canal and mandibular third molars were included. Due to methodological heterogeneity among studies, a qualitative synthesis was performed. After screening titles, abstracts, and full texts, 29 studies met the eligibility criteria, including 15 using panoramic radiographs, 9 using CBCT, and 5 evaluating both imaging modalities. Protocol registration was not performed. AI-based models demonstrated favourable diagnostic performance for identifying and classifying the spatial relationship between the mandibular canal and mandibular third molars. Reported accuracy values frequently ranged from 80% to 97%, while sensitivity and specificity commonly exceeded 85%. Area under the ROC curve (AUC) values ranged from 0.84 to 0.98 across studies, indicating strong discriminatory capacity of CNN-based models for canal segmentation, spatial classification, and risk assessment tasks using panoramic radiographs and CBCT images. Current evidence suggests that AI may support radiographic interpretation and pre-surgical assessment of the relationship between the mandibular canal and mandibular third molars. However, methodological heterogeneity and limited external validation highlight the need for further standardized multicenter studies before broader clinical implementation.
Araújo et al. (Sat,) studied this question.