Artificial intelligence in spinal fracture imaging: from clinical application to translational promise

Abstract

Artificial intelligence (AI) is reshaping medical imaging, and deep learning together with machine learning are now being used to detect, classify, and differentially diagnose spinal fractures. In this Review, we survey the principal AI architectures that currently support spinal imaging, including convolutional neural networks, object detectors, and hybrid models, and we assess their reported performance for osteoporotic, traumatic, and pathological vertebral fractures. Our synthesis indicates that AI can achieve high diagnostic accuracy in controlled, often single-centre retrospective settings, but its performance consistently declines for mild fractures, degenerated spines, and postoperative anatomy. Critically, the majority of published models lack external validation, and many studies suffer from methodological shortcomings such as selective reporting, insufficient sample sizes, and absence of predefined analysis plans—factors that systematically inflate apparent accuracy. Beyond detection, we explore the growing applications of AI in surgical planning, complication prediction, and functional outcome estimation, while emphasising that most evidence remains preliminary. We identify persistent gaps, including underuse of foundation models and public benchmarks, and limited independent evaluation of commercial tools. We conclude by proposing a shift from performance reporting to methodologically rigorous, prospectively designed studies that incorporate bias assessment, standardised reporting, and patient-centred endpoints—steps essential to translate technical promise into tangible patient benefit.