AI-Driven Decision Thresholds in Cariology: A Systematic Review of Lesion Stage Detection on Bitewing Radiographs.

Abstract

Introduction: This systematic review evaluates the stage-specific diagnostic accuracy of artificial intelligence (AI) models for caries detection and compares their performance with human examiners.

Methods: Following PRISMA 2020 guidelines, 4 databases (PubMed/Scopus/Embase/Web of Science) were searched up to October 2025. Nineteen studies using bitewing radiographs and reporting at least 1 diagnostic metric stratified by lesion stage (E₁-D₃, or equivalent) were included. Data extraction included model architecture, dataset characteristics, annotation, and stage-specific outcomes. Risk of bias was assessed using QUADAS-AI. Due to heterogeneity in staging systems (binary, ICDAS, ICCMS) and reported metrics, a narrative synthesis was conducted.

Results: AI models, especially YOLO variants, U-Net architectures, and ResNet classifiers, demonstrated consistently higher sensitivity (0.56-0.96) than human examiners (0.21-0.83) for early enamel lesions. Across the 19 included studies, 17 reported stage-specific outcomes for enamel and 18 for dentin lesions, with 10 studies providing direct AI-human comparisons. For moderate and advanced dentin lesions, AI performance was comparable to that of human examiners, with strong F1-scores (0.63-0.93), high area under the curve (0.73-0.97), and low false-negative rates, although specificity was less consistently reported. QUADAS-AI identified a high risk of bias in at least 1 domain in 12 studies due to patient selection, insufficient blinding, and limited external validation.

Conclusion: AI systems show potential as adjunctive tools for stage-specific caries detection on bitewing radiographs. Their use may be helpful for preliminary screening and flagging of possible early enamel lesions, but performance remains variable and reflects agreement with expert radiographic interpretation rather than validation against biological truth. Evidence certainty was limited by heterogeneity and methodological bias. Standardized reporting, external validation, and unified lesion depth criteria are required before clinical integration.