POROVNÁNÍ DIAGNOSTICKÉ PŘESNOSTI SYSTÉMŮ UMĚLÉ INTELIGENCE A JUNIORNÍCH RADIOLOGŮ PŘI DETEKCI FRAKTUR: RETROSPEKTIVNÍ, MULTI-READER STUDIE

Authors:
Open Access
Publisher website
Google Scholar
Abstract
A retrospective, single-center, multi-reader, multi-case study evaluated the diagnostic performance of two artificial intelligence systems (AI 1, AI 2) in detecting primary fractures on musculoskeletal X-ray images and compared it with the performance of three clinically active radiologists (RAD 1–3). A total of 1,137 anonymized images taken between March 5 and 11, 2025, were selected from the PACS archive of the Olomouc University Hospital; the prevalence of fractures was 10.4%. The reference standard was established by majority vote consensus of three independent radiologists, with substantial agreement confirmed by Fleiss‘ κ = 0.739. The power analysis required a minimum of 110 fractures and 900 negative cases; the actual sample comfortably exceeded these limits. The statistical evaluation was based on four-field contingency tables with Wilson’s 95% confidence intervals (CI); differences between readers were tested using McNemar’s test with Holm’s correction. AI 1 achieved a sensitivity of 0.941 (95% CI 0.883–0.971) and a specificity of 0.930 (0.913–0.944), AI 2 achieved a sensitivity of 0.932 (0.872–0.965) and specificity of 0.935 (0.918–0.949). The sensitivity of both algorithms did not differ statistically (p = 0.79); However, AI 2 showed higher specificity than AI 1 (p < 0.05). Radiologists achieved sensitivity of 0.915–0.941 and specificity of 0.944–0.977. Their sensitivity was comparable to both AIs, while their specificity exceeded AI 1 and was similar to or slightly higher than AI 2. The negative predictive value of all evaluated entities exceeded 0.990. Subanalyses of six anatomical regions and four age cohorts confirmed the consistently high sensitivity of both AIs, with the highest accuracy in the shoulder/clavicle region and the greatest decrease in specificity of AI 1 in the hand/wrist region. The study shows that both systems provide fracture detection with sensitivity comparable to that of the radiologists involved.