Comparative analysis of AI support levels in clinical interpretation of traumatic pelvic radiographs

Abstract

Plain pelvic radiographs (PXR) remain crucial for initial trauma assessment, yet interpretation challenges persist. While artificial intelligence (AI) shows promise, its practical impact across specialties remains unexplored. We conducted a retrospective image-based, multi-reader multi-case (MRMC) study using a standardized, prospectively planned evaluation protocol. A total of 26 physicians (8 radiologists, 10 emergency physicians, 8 trauma surgeons) interpreted 150 PXRs in three sequential sessions: without AI, with AI-alert, and with AI-visual guidance. AI assistance improved overall diagnostic accuracy from 0.870 to 0.940 (p < 0.001) and reduced interpretation time from 22.70 to 9.58 s (p < 0.001). Non-radiologists showed substantial improvements, with emergency physicians demonstrating increases in specificity (26.2%, p = 0.006) and positive predictive value (41.5%, p = 0.006). Trauma surgeons with AI-visual guidance achieved comparable accuracy to unaided radiologists (0.940 vs. 0.920, p = 0.556). Tailored AI assistance effectively bridges the performance gap between radiologists and non-radiologists while reducing reading time. These findings suggest AI integration could enhance clinical workflow efficiency across specialties in trauma care settings.

Ähnliche Arbeiten