AI assistance reduces missed fractures on musculoskeletal radiographs: A multicentre multi-reader crossover study

Purpose: To quantify the impact of AI software on fracture detection on musculoskeletal radiographs in a multicentre multi-reader crossover study. Methods: In this retrospective multicentre study, 726 radiographs from five Czech hospitals were interpreted twice by the seven clinicians: first without AI and, after a 30-day washout, with AI decision support. The endpoint was binary fracture detection. The reference standard was defined by independent review by one trauma surgeon and one radiologist; cases with discordant expert assessments were excluded. Paired analyses were performed on the expert-agreement set. Sensitivity, specificity, positive likelihood ratio (PLR), and negative likelihood ratio (NLR) were calculated per reader, and overall performance was summarised using a nonparametric case-level bootstrap. Results: The primary analysis set included 630 radiographs, of which 197 were fracture-positive and 433 were fracture-negative. Overall bootstrap-estimated sensitivity was 86.26% (80.82;90.41) without AI and 94.56% (90.54; 97.00) with AI ( p < 0 . 001 ). Overall specificity was 93.79% (91.14;95.72) without AI and 90.21% (87.08;92.69) with AI ( p = 0 . 035 ). Overall NLR improved from 0.15 (0.10;0.21) to 0.06 (0.03;0.11) ( p = 0 . 008 ), while overall PLR decreased from 14.24 (9.94;20.84) to 9.75 (7.36;12.97) ( p = 0 . 134 ). Sensitivity increased in six of seven readers, and the largest gain occurred in the reader with the lowest baseline sensitivity. Conclusion: AI assistance reduced missed fractures across readers, with a modest and reader-dependent increase in false-positive fracture calls. These findings suggest that AI-assisted interpretation may function as a useful second-reader aid in musculoskeletal radiography.