AI-Based Detection of Proximal Caries Across Radiographic Modalities: A Systematic Review and Meta-Analysis.
TL;DR
AI models show promising diagnostic accuracy for proximal caries, with bitewing radiographs yielding slightly better performance than panoramic views, and specific sources of performance heterogeneity that have previously confounded pooled estimates in the field are identified.
Abstract
Objectives
To evaluate diagnostic performance of artificial intelligence (AI) models in detecting proximal caries across different radiographic modalities.
Methods
We systematically searched in five electronic databases: Web of Science, PubMed, IEEE Xplore, ScienceDirect and CNKI. Essential study characteristics, AI models and their accuracy metrics for proximal caries detection were extracted. Methodological quality of eligible studies was assessed with QUADAS-2. Studies judged to be of adequate quality were retained for meta-analysis.
Results
Twenty studies were included: fifteen employed bitewing radiographs, three utilized panoramic radiographs, and two adopted periapical radiographs. AI accuracy ranged from 28.5% to 100%. Due to the unavailability of essential 2 × 2 contingency data, only ten studies (six bitewing, three panoramic, and one periapical) were eligible for meta-analysis, yielding a pooled sensitivity of 76% (95% CI: 70%-80%), specificity of 94% (95% CI: 90%-96%), and Summary Receiver Operating Characteristic (SROC) AUC of 0.90 (95% CI: 0.87-0.92). Substantial heterogeneity was observed.
Conclusion
AI models show promising diagnostic accuracy for proximal caries, with bitewing radiographs yielding slightly better performance than panoramic views. While AI holds potential as a clinical decision-support tool, high heterogeneity and limited external validation remain significant barriers to clinical translation. Future work should prioritize prospective, multicenter validation and open datasets to support clinical translation. ADVANCES IN KNOWLEDGE This systematic review covers studies up to October 2025, simultaneously evaluate AI performance of different modalities, enabling direct cross-modal comparison. By stratifying analyses by task type, dataset size, and imaging modality, this study identifies specific sources of performance heterogeneity that have previously confounded pooled estimates in the field. This review also establishes evidence thresholds for clinical implementation, serving as a methodological checkpoint to guide future research.