YOLO-based AI boosts radiologist fracture detection accuracy and slashes read time in multi-reader study

AI-assisted appendicular fracture detection (AUC 0.847, 95% CI 0.811–0.881) improved accuracy for 2 of 3 radiologists (~76% → 83%, p≤0.001) and cut median read time 33% (12.9→8.6 s) in a retrospective MRMC study; prospective multicenter validation still needed.

• Retrospective MRMC study: YOLO26x model trained on 8,690 high-resolution (1280×1280 px) appendicular radiographs; tested on 500 balanced cases (250 fracture-positive, 250 negative) across institutional and open-source data — AUC consistent across both subsets (0.843 vs. 0.852).
• Inter-reader agreement improved substantially with AI assistance (Fleiss' κ 0.432 → 0.642); sensitivity gains ranged 2.4–16.4 percentage points; beneficial decision changes outnumbered detrimental ones 197 vs. 104 across 1,500 reader-case pairs.
• Key limitation: only 3 radiologists in the MRMC component and a sequential (unassisted-then-assisted) rather than randomized crossover design introduces potential learning/fatigue bias; no prospective or multicenter external validation performed.

RadPigeon summaries are original and for information only. They are not clinical advice.