Validation of Artificial Intelligence in Detecting Acute Cervical Spine Fractures on CT.
Authors
Affiliations (2)
Affiliations (2)
- From the Department of Radiology (M.H.D.H., N.T.M., K.C., N.V., A.B., N.P., C.W., J.M., A.B., M.M.-B.),University of Washington, Seattle, WA, USA and Department of Radiology (M.H.D.H., M.M.-B.), University of Alabama-Birmingham, Birmingham, AL, USA.
- From the Department of Radiology (M.H.D.H., N.T.M., K.C., N.V., A.B., N.P., C.W., J.M., A.B., M.M.-B.),University of Washington, Seattle, WA, USA and Department of Radiology (M.H.D.H., M.M.-B.), University of Alabama-Birmingham, Birmingham, AL, USA. [email protected].
Abstract
Cervical spine fractures require prompt and accurate diagnosis to minimize risk of long-term neurological impairment. Artificial intelligence (AI) algorithms could assist in the detection of cervical spine fractures. This study aims to evaluate the diagnostic accuracy of AI, specifically Aidoc cervical spine tool, in detecting acute cervical spine fractures on CT, in comparison to radiologists. A retrospective study was performed between November 2020 and March 2021 of CT studies that included the entire cervical spine from the UW Medicine database; these included CT cervical spine, CT full spine, CT neck and CTA neck studies. Exams were evaluated for the presence of acute fractures using the Aidoc tool and compared to radiologist reports, with a radiologist consensus panel serving as the reference standard when AI-radiologist discordance arose. Parameters such as location of fracture (upper vs lower cervical spine, anterior vs posterior elements), performance in different environments, between dedicated spine CT scans and CTA/CT neck., and the presence of hardware were also collected for subgroup analyses. 2652 CT exams were included in the study i.e. 1180 Cervical spine CT, 467 Full spine CT, 474 Neck CTA, and 531 Soft tissue Neck CT. There were 110 acute cervical spine fractures in our cohort (prevalence of 4.1%); AI correctly identified 98 fractures but with 21 false positives. AI had a sensitivity and specificity of 89.1% (95% CI=81.7-94.2%) and 99.2% (95% CI=98.7-99.5%), while radiologists had a sensitivity and specificity of 94.5% (95% CI=88.5-98.0%) and 99.7% (95% CI=99.4-99.9%), respectively. In a simultaneous omnibus statistical test for both sensitivity and specificity, radiologist interpretation outperformed AI interpretation (p = 0.02). Reasons for AI false positives in our cohort included atherosclerotic calcifications, chronic fractures, beam artifact, osteophytes and unfused apophysis misinterpreted as acute fractures. Our study demonstrated that radiologists outperformed AI in detection of acute cervical spine fractures on CT.