Artificial Intelligence in the Imaging-Based Analysis of Third Molars: A Systematic Review.

BACKGROUND: This systematic review synthesizes the current evidence regarding the application of artificial intelligence (AI) in third molar imaging analysis, highlighting current evidence, methodological limitations, and challenges for future clinical translation. PURPOSE: To evaluate the effectiveness of AI-based methods applied to dental imaging for the detection, classification, and surgical planning of third molars. STUDY SELECTION: Electronic searches were conducted in PubMed, Scopus, and Web of Science to identify studies that developed, validated, or evaluated AI methods for third molar analysis using dental imaging. Data extraction included study objectives, model architectures, imaging modalities, dataset characteristics, target tasks, validation strategies, and performance metrics. Two independent reviewers performed the data extraction, and disagreements were resolved through consensus with a third reviewer. Risk of bias and applicability concerns were assessed using the QUADAS-2 tool. RESULTS: Of 227 identified records, 32 studies met the eligibility criteria. Most studies evaluated convolutional neural network-based models, particularly You Only Look Once architectures, using panoramic radiography as the primary imaging modality. The main evaluated tasks included third molar detection, segmentation, classification, and surgical difficulty prediction. Performance varied according to model architecture, target outcome, and validation strategy, with greater variability observed in complex tasks such as root morphology assessment, angular measurements, and surgical difficulty prediction. Overall, evidence was limited by methodological heterogeneity, inconsistent reporting of performance metrics, and limited external validation. CONCLUSIONS AND RELEVANCE: AI-based methods show promising reported performance in the detection and classification of third molars using dental imaging. However, more complex tasks still require methodological refinement and external validation before their potential integration into clinical workflows can be adequately assessed.