Artificial Intelligence for Radiographic Assessment and Severity Grading of Peri-Implant Marginal Bone Loss: A...

To systematically evaluate the diagnostic performance, methodological quality, reporting transparency, and descriptive translational staging of Artificial Intelligence (AI) models applied to dental imaging for the radiographic assessment and severity grading of peri-implant marginal bone loss.

The review was conducted according to PRISMA 2020 and PRISMA-DTA. PubMed (MEDLINE), Scopus, Web of Science, and Embase were searched without date restrictions. Ten studies met the eligibility criteria.

Nine studies used retrospective single-center designs, and one included multicenter data without independent external validation. AI tasks included landmark detection, quantitative measurement, segmentation, bone-loss detection, and severity grading. Reported performance was generally favorable but highly task-specific. Implant detection reached an average precision of approximately 0.99 in one study, whereas average precision for marginal bone-loss lesion detection was 0.47. Implant segmentation achieved a Dice similarity coefficient of 0.986, while reported F1-scores reached 97.1%. Sensitivity and specificity were reported in only a subset of diagnostic evaluations and varied according to the unit of analysis. Quantitatively interpretable measurement-error reporting was limited; only one study reported deviations that could be converted to approximately 0.17 mm using the study-specific image calibration. No study provided a complete set of clinically interpretable absolute marginal bone-loss measurement-agreement outcomes in millimetres. Substantial heterogeneity was observed in annotation protocols, validation strategies, dataset partitioning, task definitions, and performance reporting. No study performed prospective evaluation or independent external multicenter validation.

Current AI systems show promising technical capability for radiographic assessment of peri-implant marginal bone loss. However, the evidence remains limited by retrospective study designs, incomplete reproducibility, lack of external validation, and uncertain clinical generalizability.

The clinical utility of AI as an adjunctive tool for radiographic interpretation or longitudinal monitoring was not tested in the included studies and remains to be established through prospective evaluation, external validation, and direct assessment of clinician-AI interaction.