AI detection of peri-implantitis on 2D radiographs: a systematic review and diagnostic accuracy assessment.

Peri-implantitis is a major biological complication that compromises the longevity of dental implants, and accurate radiographic assessment is essential for early detection and intervention. Artificial intelligence (AI) has shown promising performance in dental imaging; however, the evidence supporting its diagnostic utility for peri-implant disease on two-dimensional radiographs remains fragmented. To systematically evaluate and synthesise the performance, methodological quality, and certainty of evidence of deep-learning models developed for the detection or assessment of peri-implantitis and peri-implant bone loss using two-dimensional dental radiographs. A systematic search of PubMed, Scopus (including Embase-indexed records), and the Cochrane Library was conducted from inception to January 2025. Studies using deep-learning models to detect peri-implantitis, quantify peri-implant bone loss, or classify peri-implant defect morphology on two-dimensional radiographs were eligible. Screening, data extraction, and risk-of-bias assessment (QUADAS-2) were performed in duplicate. Due to substantial heterogeneity in outcomes, diagnostic definitions, and reporting formats, a formal meta-analysis was not feasible; results were synthesised narratively, with descriptive pooling where appropriate. Certainty of evidence was evaluated using GRADE for diagnostic accuracy. Twenty-eight records were screened, and eight studies met the inclusion criteria. All were retrospective, single-centre investigations published between 2021 and 2025. Deep-learning models demonstrated high technical performance within internal datasets, with binary classifiers achieving sensitivities up to 0.90-0.98, specificities up to 0.95, and segmentation models yielding Dice coefficients exceeding 0.97. Multi-class and measurement-based systems also showed strong agreement with clinician assessments. However, all studies exhibited high risk of bias in patient selection and index-test domains, relied exclusively on internal validation, and used heterogeneous diagnostic targets and reference standards. Only two studies reported sufficient information to derive sensitivity and specificity. The overall certainty of evidence was judged to be very low. Deep-learning systems show strong technical promise for identifying peri-implant bone loss and peri-implantitis on two-dimensional radiographs, frequently matching or surpassing clinician-level performance within controlled settings. However, serious methodological limitations and the absence of external validation constrain the certainty and generalisability of current evidence. AI-based peri-implant diagnostic systems should therefore be considered investigational, and robust multicentre prospective studies are required before clinical implementation.