To evaluate the caries detection performance of a commercially available AI-system (Nostic) and human raters for radiographic caries detection using tooth sections as reference. Radiographs and corresponding tooth sections from extracted teeth (548 approximal and 91 cervical sites) were assessed by three dentists and AI-system (graded/probabilistic outputs). Performance versus reference was quantified using ROC/PR analyses, calibration metrics, and threshold-based measures with bootstrapped confidence intervals. Rater-AI-system agreement/disagreement subsets were analysed. For early dentine lesions (D3), four decision approaches (rater-only, AI-system-only, rater-OR-AI-system, rater-AND-AI-system) were compared using a weighted clinical-loss analysis across varying false-positive penalties. For approximal enamel lesions (D1/2), AI-system was more sensitive than raters (0.60 vs 0.42) with slightly lower specificity (0.80 vs 0.84) and modest discrimination (AUC 0.70 vs 0.64). For approximal dentine lesions (D3/4), discrimination was high (AUC 0.90 AI-system vs 0.86 rater); The AI-system was more sensitive (0.84 vs 0.72) while raters were more specific (0.94 vs 0.88). For cervical dentine lesions (D3/4), both performed well (AI-system: Se/Sp 0.92/0.84; rater: 0.97/0.77; AUC 0.91 vs 0.87). In D3 decision strategies, rater-only prioritised specificity (Se/Sp 0.60/0.85), AI-system-only prioritised sensitivity (0.75/0.77), OR reduced false negatives (0.84/0.73), and AND reduced false positives (0.51/0.89), with AND yielding the lowest clinical loss at higher false-positive penalties. The AI-system provides complementary information that becomes clinically relevant when integrated into structured human-AI-system decision rules. Context-dependent use may support minimally invasive caries management. Combining human assessment with AI-system may improve preventive and operative decision-making by balancing false negatives and false positives.
Ganß et al. (Sun,) studied this question.