In a real‑world test of 1,672 X‑rays, a commercial AI tool nudged junior doctors’ accuracy from 87% to 90%, but it also added some false alarms—especially around growth plates.  

What’s new

Researchers at a German children’s hospital evaluated an approved fracture‑detection program on everyday emergency‑department X‑rays. The AI performed well on its own and, when shown to inexperienced physicians after their first read, slightly improved their ability to tell which children had a fracture. The study appears in European Radiology. The graphic on page 2 shows the small but measurable bump in reader accuracy with AI.  

Key numbers (plain language)

• Who was studied: 1,657 children (median age ~11; 59% boys) with 1,672 radiograph studies from a single pediatric ED. Nearly half of studies had a fracture. These details are listed in Table 1 on page 4.  
• AI alone: Sensitivity 92%, specificity 83%, accuracy 87% across body regions (see Table 2 and Figure 3 on page 5).  
• Helping junior doctors: With AI, reader sensitivity rose from 83.7% → 87.3%, specificity from 90.7% → 92.4%, and overall accuracy from 87.3% → 89.9%—a net +2.6 percentage‑point gain. However, in ~1.5–2% of cases the AI nudged a correct diagnosis to a wrong one (Figure 5 on page 7).  
• High‑stakes fracture types: The AI caught 100% of subtle proximal tibia fractures and 96% of medial malleolus fractures, but only 68% of tricky radial condyle (elbow) fractures (Table 4 and images on page 6).  

Where AI struggled

Growth plates and normal bone variants often fooled the software. The example image on page 4 shows the AI falsely flagging parts of a child’s foot; the false‑positive list in Table 3 (page 5) highlights the fifth metatarsal apophysis and toes as frequent pitfalls. Most false positives were labeled “low confidence” by the tool—useful context for clinicians.  

Why it matters

Reading children’s X‑rays is hard—bones are still growing and fractures can be subtle. An AI “second set of eyes” can reduce missed injuries and support training for new clinicians. The take‑home from this study: AI can improve safety a little, but it doesn’t replace expert review, and performance varies by body part (elbows remain challenging).  

Caveats

This was a single‑center, retrospective study using one software product. The authors note that cost‑vs‑benefit needs careful thought, and that results could differ with software updates or in other hospitals. Future studies should test effects on follow‑up visits, complications, and overall costs.  

Bottom line

In a busy pediatric ER, AI fracture detection is a helpful assistant—not a final judge. Used thoughtfully, it can catch a few more breaks and give clinicians added confidence, while reminding everyone that growing bones still demand human expertise.  

Source: Ziegner M., Pape J., Lacher M., et al. “Real‑life benefit of artificial intelligence–based fracture detection in a pediatric emergency department,” European Radiology* (2025). See pages 2, 4–7 for figures and tables cited above.* 

Ziegner, M., Pape, J., Lacher, M., Brandau, A., Kelety, T., Mayer, S., Hirsch, F. W., Rosolowski, M., & Gräfe, D. (2025). Real-life benefit of artificial intelligence-based fracture detection in a pediatric emergency department. European radiology, 35(10), 5881–5890. https://doi.org/10.1007/s00330-025-11554-9

Editor’s note: This article is for information only and is not a substitute for professional medical advice.