DEEP LEARNING–ASSISTED FRACTURE DETECTION ON PLAIN RADIOGRAPHS: DIAGNOSTIC ACCURACY, CLINICAL INTEGRATION, AND FUTURE DIRECTIONS

Introduction and Purpose. Fractures are a major global health burden, and a non-trivial fraction of fractures presenting to emergency and urgent care settings are initially missed on plain radiographs, causing diagnostic delay and avoidable morbidity. This review appraises the current state of deep learning for fracture detection on conventional radiography, addressing standalone accuracy, the effect of algorithmic assistance on readers, and the conditions of clinical integration. Brief Description of the State of Knowledge. Across recent meta-analyses, pooled sensitivity and specificity for fracture detection on radiographs cluster in the low-to-mid nineties, with summary AUROC approaching 0.97 in the largest synthesis. The literature is unevenly distributed across the skeleton: the wrist and the hip dominate the evidence base, the pediatric elbow is moderately developed, and other regions remain represented by only a handful of studies. Algorithmic assistance lifts reader sensitivity by several percentage points without a specificity penalty, with the largest gains in clinicians who have less musculoskeletal imaging experience. Summary. Deep learning is mature for narrow fracture-detection tasks in the appendicular skeleton, but not yet for broad deployment claims. External validation across institutions and vendors, prospective evaluation tied to clinical endpoints, and standards-based integration into PACS and reporting infrastructure are the conditions that would move these tools from accuracy benchmarks toward measurable patient benefit. The clearest near-term value lies in decision support that narrows the diagnostic gap for less experienced readers.