Abstract
Background: Peri-implantitis is a common implant complication requiring early detection to prevent bone loss and implant failure. Deep learning models show promise for enhancing radiographic diagnosis.
Objectives: This review systematically evaluated the diagnostic performance of deep learning models in detecting peri-implant marginal bone loss on radiographic images.
Materials and methods: A comprehensive search of PubMed, EMBASE, CENTRAL, ClinicalTrials.gov, and ProQuest identified studies published between 2010 and July 2025. Two reviewers independently screened studies, extracted data, and assessed methodological quality using QUADAS-2. Diagnostic metrics, including sensitivity, specificity, F1-score, area under the curve (AUC), were synthesized using random-effects meta-analysis. Heterogeneity and publication bias were evaluated using I2 statistics, meta-regression, and funnel plots.
Results: Five studies comprising 12,545 periapical and panoramic radiographs met inclusion criteria. Deep learning models achieved pooled sensitivity of 88%, specificity of 91%, and AUC of 0.95, indicating high diagnostic performance. Positive and negative likelihood ratios suggested strong clinical utility. Quality was generally good, though reporting of implant characteristics and data augmentation was inconsistent. Meta-regression revealed that dataset size and unit of analysis influenced accuracy, whereas imaging type did not. No publication bias was found.
Conclusion: Deep learning models demonstrate high accuracy in detecting radiographic marginal bone loss, potentially indicating peri-implantitis but cannot substitute for comprehensive clinical assessment.
Clinical relevance: These models offer a promising adjunct for radiographic detection of marginal bone loss, supporting clinicians in early diagnosis and timely interventions.