Comparative evaluation of proximal caries detection methods using human, artificial intelligence, and micro-CT...

To validate the diagnostic accuracy of ICDAS visual examination, conventional bitewing radiography, and artificial intelligence (AI)-assisted bitewing analysis (Overjet Caries Assist) for classifying proximal caries lesion depth using micro-computed tomography (micro-CT) as the reference standard. Specifically, this study evaluated: (1) reliability and inter-method agreement with micro-CT; (2) agreement between AI-assisted and human bitewing interpretation; and (3) diagnostic performance at enamel and dentin thresholds relevant to preventive and restorative decision-making.

Twenty-seven extracted human teeth representing ICDAS scores 0-6 were imaged using bitewing radiography and micro-CT. Two calibrated raters independently scored bitewing and micro-CT images using an ICDAS-analogous scale (E0-D3). AI-assisted scores were generated from the same bitewing images. Scores were categorized into four groups: sound, enamel, outer dentin, and inner dentin. Micro-CT consensus served as the reference standard. Intra-rater reliability was assessed using weighted Cohen's kappa. Inter-method agreement was evaluated using weighted kappa and Spearman's rho. Diagnostic accuracy (sensitivity and specificity) was calculated at enamel and dentin thresholds with 95% confidence intervals. Bitewing and AI performance were compared using a one-sided Wilcoxon signed-rank test (α = 0.05).

Intra-rater agreement was high across modalities (κ = 0.93-0.94). Agreement with micro-CT was substantial for ICDAS (κ = 0.67) and bitewing (κ = 0.70), and moderate for AI (κ = 0.53). Bitewing classified lesion depth significantly more accurately than AI (p = 0.042). At the enamel threshold, bitewing and AI demonstrated 100% sensitivity, whereas ICDAS achieved 81%. At the dentin threshold, ICDAS showed 100% sensitivity and specificity, while bitewing and AI achieved sensitivities of 75% and 56%, respectively. AI and bitewing errors were exclusively under-classifications.

No single method was superior across all diagnostic thresholds. AI-assisted and conventional bitewing imaging supported enamel-level detection, whereas ICDAS provided superior identification of dentin involvement. AI should be considered a screening adjunct rather than a standalone diagnostic method, supporting a multimodal diagnostic approach integrating visual and radiographic assessment.

AI-assisted caries detection reliably identifies proximal enamel lesions, supporting its use as a screening adjunct in preventive care workflows. However, AI demonstrated reduced sensitivity for dentin involvement and systematically underestimated lesion depth. Clinical visual examination remains essential for accurate diagnosis and treatment planning, reinforcing the value of a multimodal diagnostic approach integrating visual and radiographic assessment.