Title : Artificial intelligence for detection of trauma related lower limb fractures on plain radiographs
Abstract:
Objective: Lower limb fractures due to trauma place a heavy burden on emergency departments; however, there is known reader variability in their detection. Deep learning offers a potential solution for reducing variability. This review evaluates the diagnostic accuracy of deep learning algorithms for the detection of traumatic lower limb fractures on plain radiographs.
Methods: This PRISMA 2020-compliant systematic review and meta-analysis was registered with (PROSPERO: CRD420261353635). MEDLINE, Embase, and the Cochrane Library were searched from inception to December 2025. Studies evaluating Artificial Intelligence (AI) applied to plain radiographs for lower-limb fracture detection were eligible and methodological quality was assessed using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool. A bivariate random-effects model was used to pool sensitivity and specificity . A Summary Receiver Operating Characteristic (SROC) curve was also generated. Deeks' funnel plot assessed publication bias . Subgroup analyses evaluated adult versus paediatric populations and internally versus externally validated models.
Results: Of 1,203 identified records, 24 studies met eligibility criteria (23,594 radiographs; 13,457 depicting fractures). Pooled sensitivity was 85.6% (95% confidence interval CI: 81.8–88.7%; I²= 89.7%) and pooled specificity was 94.1% (95% CI: 90.5–96.3 %; I²= 94.9%). The area under the SROC curve was 0.938 (95% CI: 0.906–0.962). Deeks' funnel plot indicated low publication bias (p=0.272). Internally validated models had significantly higher sensitivity than externally validated models (90.3% vs 83.5%; p=0.047).
Conclusion: Deep learning algorithms achieve high diagnostic accuracy in detecting lower-limb fractures on plain radiographs. Substantial between-study heterogeneity and a clinically meaningful sensitivity reduction in externally validated models indicate that prospective multicentre validation is required before routine clinical deployment.
