Abstract
Introduction: Convolutional neural networks (CNNs) and transformers are artificial intelligence (AI) models used for accurate periodontal bone loss (PBL) diagnosis. This study compared the diagnostic accuracy between CNN and transformer models in detecting PBL.
Methods: We searched ten databases: PubMed, Scopus, Cochrane, Embase, WorldCat, Lens, Google Scholar, Taylor & Francis, ScienceDirect, and ClinicalTrials.gov from 2015 to 2025 to identify studies that used CNN and/or transformer models on human dental radiographs, evaluating diagnostic performance based on area under the curve, accuracy, sensitivity, specificity, correlation matrix, F1-score, precision, and recall. Risk of bias was assessed using PROBAST+AI. Meta-analyses using random-effects models estimated pooled diagnostic odds ratio (DOR), sensitivity, and specificity.
Results: Seventeen studies met the eligibility criteria: 14 evaluated CNN-based models, and 3 evaluated transformer-based models, comprising a total of 26,502 cases and 55,076 controls. The overall risk of bias was low. CNN models demonstrated pooled sensitivity and specificity of 0.85 (95% CI: 0.77–0.90) and 0.88 (95% CI: 0.80–0.93), with a DOR of 41.69 (95% CI: 19.41–89.52). Transformer models showed a sensitivity of 0.87 (95% CI: 0.74–0.94), specificity of 0.80 (95% CI: 0.34–0.97), and DOR of 25.64 (95% CI: 3.47–189.63). Heterogeneity was substantial (I²>98%), with radiograph type and study design identified as significant contributors. CNN results remained robust across sensitivity analyses, while transformer estimates were more sensitive to individual studies.
Conclusion: Current evidence demonstrates greater diagnostic accuracy of CNN-based models for periodontal bone loss detection, while comparative conclusions for transformer-based models remain premature.