Rethinking diagnostic performance metrics in imaging AI: lessons from caries detection on radiographs

Artificial intelligence (AI) systems for radiographic caries detection are commonly evaluated using a small set of performance metrics, yet these measures are frequently reported and interpreted without the contextual information required for clinically meaningful conclusions. This narrative review focuses on performance metrics - confusion-matrix measures (sensitivity, specificity, accuracy, predictive values, F1 score and related summaries) and discrimination metrics (ROC-AUC and PR-AUC) -and discusses how their interpretation depends on explicit reporting of the disease definition (grading cut-offs), unit of analysis (tooth surface/patient), decision threshold(s), and prevalence/case-mix. We additionally cover complementary metric-based assessments, including calibration (calibration curves, expected calibration error, Brier score), decision-analytic metrics (decision curve analysis and cost-sensitive clinical loss), and robustness summaries (uncertainty and subgroup/worst-group performance). A structured summary is provided of formulas, interpretive limitations, and the minimum information that should accompany each metric to support transparent reporting, comparability, and transportability.