Deep learning segmentation of the mandibular canal on cone-beam computed tomography: a systematic review

Introduction: The mandibular canal houses the inferior alveolar nerve, and its accurate localization on cone-beam computed tomography is essential to avoid nerve injury during third molar surgery and implant placement. We synthesized the accuracy and validation practices of deep-learning models that segment the mandibular canal on cone-beam computed tomography. Methods: We searched PubMed, OpenAlex, and Crossref for studies published from 2013 to 2025, screened 245 records after deduplication, and included studies that evaluated a deep-learning model for mandibular canal segmentation against manual annotation. Segmentation accuracy and reporting practices were extracted. Results: Sixteen studies were included, published between 2020 and 2025 across nine countries. Eleven studies reported the Dice similarity coefficient, which ranged from 0.55 to 0.95, with the highest values above 0.85 and the lowest below 0.6. Accuracy varied widely with dataset, annotation protocol, and the inclusion of anatomical variants such as bifid canals and the anterior loop. Few studies used external validation, and public datasets were the exception rather than the rule. Conclusions: Deep-learning models can segment the mandibular canal on cone-beam computed tomography with Dice values that approach expert annotation, but the evidence is limited by small and mostly private datasets, heterogeneous reporting, and scarce external validation, so clinical deployment still requires prospective multicentre evaluation.