Surgical removal of impacted mandibular third molars (M3M) is a common procedure, yet preoperative difficulty assessment remains heterogeneous and operator-dependent. While artificial intelligence (AI) has been increasingly investigated to automate radiographic interpretation and risk prediction, the literature lacks a comprehensive synthesis specifically focusing on how AI operationalizes surgical difficulty. Following PRISMA-ScR guidelines, this scoping review identified 12 original studies from PubMed, ScienceDirect, and Google Scholar up to December 31, 2025. Results show that most models utilize panoramic radiographs as primary input, with architectures evolving from traditional CNNs to advanced Transformers and YOLO-based detectors. We found that surgical difficulty is operationalized through non-equivalent endpoints, ranging from subjective radiographic indices to objective intraoperative time. Our analysis further highlights a paradigm shift where Transformer-based models outperform CNNs by capturing the long-range spatial dependencies essential for complex tooth-nerve risk assessment. Critical appraisal via PROBAST and CLAIM reveals that while models are highly applicable, 100% carry a high risk of bias due to insufficient external validation and reporting deficits. We propose the M3M-AI Framework to harmonize input fidelity, core outcome sets, and model explainability, providing a roadmap for reliable clinical translation and robust cross-study comparison.
Nguyen et al. (Wed,) studied this question.