Radiographers' accuracy in interpreting acute musculoskeletal X-rays when supported by artificial intelligence - A multi-site cross-sectional study.
Authors
Affiliations (3)
Affiliations (3)
- Department of Optometry, Radiography and Lighting Design, University of South-Eastern Norway, P.O. Box 4, 3199, Borre, Norway. Electronic address: [email protected].
- Department of Optometry, Radiography and Lighting Design, University of South-Eastern Norway, P.O. Box 4, 3199, Borre, Norway; Medical Diagnostics Clinic, Vestre Viken Health Trust, P.O. Box 800, 3004, Drammen, Norway.
- Medical Diagnostics Clinic, Vestre Viken Health Trust, P.O. Box 800, 3004, Drammen, Norway.
Abstract
The artificial intelligence (AI) tool BoneView™ has been introduced into clinical practice to support skeletal X-ray interpretation. This created a new workflow in which radiographers are expected to evaluate and act on AI output. This study assessed the accuracy of AI-supported radiographers in detecting skeletal injuries on adult trauma X-rays. In this cross-sectional study, 10 AI-supported diagnostic radiographers from 4 hospitals retrospectively assessed 542 acute musculoskeletal X-ray examinations. The radiographers interpreted the examinations with access to BoneView™ output and marked each case as either sure/unsure positive or sure/unsure negative, with an optional free-text field for comments. Sensitivity and specificity were calculated for BoneView™ and for each AI-supported radiographer, using a quality-assured radiologist's report as the reference standard. Differences in diagnostic performance between AI and AI-supported radiographers, and between AI-supported radiographers, were examined using generalized linear mixed models. AI-supported radiographers had an overall sensitivity of 94% and specificity of 86%, compared with 96% and 71%, respectively, for AI alone. The difference in diagnostic performance was primarily driven by AI-supported radiographers' higher specificity in interpreting X-rays of the pelvic/hip and foot/toe regions, but inter-radiographer variability was substantial. Radiographers with ≥5 years of experience had higher sensitivity (p 0.01) and lower specificity (p 0.02) than those with <5 years of experience. The involvement of radiographers had limited impact on sensitivity compared with AI alone but was associated with improved specificity. Variability in performance and suboptimal specificity indicate potential for further improvement through training, workflow optimization, or refinement of AI support. Understanding radiographers' diagnostic accuracy when using AI helps clarify how such tools may support triage, reduce diagnostic delays, and distribute workload in acute skeletal imaging.