Diagnostic Performance and Agreement Between Dental Students and an Artificial Intelligence System in Panoramic Radiographs

OBJECTIVES: This study aimed to compare the diagnostic behaviour and detection patterns of fourth- and fifth-year dental students with those of an artificial intelligence (AI) system in identifying dental caries and restorations on panoramic radiographs within an educational framework. METHODS: This retrospective cross-sectional study included 60 anonymised panoramic radiographs comprising 1 920 individual tooth units. Diagnostic assessments were performed independently by 60 fourth-year dental students, 60 fifth-year dental students, and an artificial intelligence (AI) system. For each tooth, the presence or absence of caries and restorations was recorded using a standardised assessment protocol without additional calibration to reflect routine educational conditions. The AI system was used under default settings without modification of internal detection parameters. Therefore, its outputs were interpreted as threshold-dependent diagnostic behaviour specific to the evaluated system rather than as a direct measure of diagnostic accuracy. A consensus-based reference derived from the majority agreement of human observers was used only for contextual comparison. Statistical analyses included Kruskal-Wallis and Mann-Whitney U tests, Cohen's kappa coefficients, and Pearson's correlation analysis (α = 0.05). RESULTS: Fourth- and fifth-year students demonstrated comparable detection patterns for both caries and restorations, whereas the AI system consistently reported fewer detections per panoramic radiograph. Intergroup agreement between student groups was moderate for caries detection and higher for restoration detection, while agreement between human evaluators and the AI system was lower. The lower detection frequency of the AI system may reflect a more restrictive diagnostic threshold rather than reduced diagnostic capability. Correlation analyses revealed weak to moderate associations among evaluator groups, indicating differences in diagnostic thresholds rather than absolute diagnostic behaviour. CONCLUSIONS: Within the limitations of this study, the AI system demonstrated a distinct diagnostic behaviour pattern compared with dental students. These findings suggest that AI may serve as a complementary tool in dental education; however, its impact on learning outcomes requires further investigation through longitudinal and interventional studies.