Can artificial intelligence diagnose the invisible? A scenario-based evaluation of AI performance in cracked tooth...

This study aimed to assess the diagnostic performance of contemporary artificial intelligence (AI) models in structured diagnostic tasks for crack-associated dental conditions, using standardized clinical scenarios without radiographic input. A scenario-based cross-sectional design was conducted using 50 standardized cases representing cracked tooth syndrome, vertical root fractures, split teeth, enamel-dentin cracks, and non-cracked control conditions. Each scenario included detailed clinical findings, including pain characteristics, pulp sensibility responses, periodontal probing, and transillumination results. AI models were evaluated using four structured multiple-choice questions addressing crack detection, diagnostic classification, pulpal diagnosis, and treatment planning. Responses were compared with expert-defined reference standards established through independent assessment and consensus agreement between two experienced endodontists (Cohen's κ = 0.92). Diagnostic performance was evaluated using sensitivity, specificity, accuracy, and Youden's index, while inter-model differences were analyzed using the Kruskal-Wallis test, followed by Bonferroni-adjusted pairwise comparisons. AI models demonstrated strong performance in structured diagnostic tasks, particularly in crack detection and diagnostic classification. Some models achieved perfect sensitivity and specificity under standardized conditions, whereas lower sensitivity was observed in diagnostically ambiguous scenarios. Performance declined in more complex tasks, particularly pulpal diagnosis and treatment planning. Higher accuracy was observed in well-defined conditions, such as vertical root fractures and split teeth, whereas lower accuracy was observed in enamel-dentin cracks and in control cases. Although the evaluated AI systems demonstrated promising performance under standardized scenario-based conditions, the absence of radiographic interpretation, patient interaction, and real-world clinical variability limits direct extrapolation of the findings to routine clinical practice.