Artificial Intelligence for Fracture Detection: Current Methods, Diagnostic Performance, and Clinical Use

Abstract

Fractures are still overlooked on initial imaging, particularly when findings are subtle or when radiographs are interpreted before specialist review. Deep-learning systems have therefore been developed to support fracture recognition on radiographs and computed tomography, with a smaller body of work addressing magnetic resonance imaging. This narrative review examines the principal model types, the datasets used for training and validation, and the performance reported across common anatomical regions. It also considers how these systems are being introduced into clinical practice. On curated test sets, many algorithms achieve diagnostic accuracy close to that of clinicians. The more clinically persuasive evidence, however, comes from assisted-reading studies, in which software improves sensitivity while largely preserving specificity. Results remain difficult to compare because datasets, reference standards, case prevalence, and validation methods vary widely. External validation, demographic fairness, automation bias, workflow design, cost, and professional accountability therefore remain central concerns. At present, fracture-detection AI is best regarded as decision support for triage or concurrent interpretation rather than an independent diagnostic service.