A French prospective study assessed two AI tools: BoneView (Gleamer) and AZtrauma (AZmed) for detecting fractures on pediatric limb radiographs. The study analyzed over 2,000 X-rays from 1,225 children, covering five regions: hand/wrist, elbow, shoulder, knee/leg, and foot/ankle. AI results were compared with radiologist readings, using CT or follow-up imaging when needed.
Both algorithms performed well overall, with higher accuracy in upper-limb imaging and reduced performance in the foot and ankle. Sensitivity dropped for subtle, non-displaced, and growth-plate–related fractures, particularly in younger children. Neither system reliably detected dislocations or soft-tissue findings.
The authors conclude that while these AI tools offer useful support in pediatric trauma radiography, they are not reliable enough for stand-alone use and require radiologist oversight. More diverse pediatric training data are needed to improve performance.
Clinical evaluation of two artificial intelligence algorithms in standard radiography for post-traumatic exploration of peripheral limbs in children
Pediatric Radiology, 2025
Abstract
Background
Pediatric traumatology represents a significant portion of emergency department visits, and advances in artificial intelligence (AI) provide promising tools for fracture detection on radiographs. However, few studies have assessed these tools on the same pediatric population.
Objective
To evaluate the diagnostic performance in a pediatric population of two commercially available AI software programs for the detection of peripheral limb fractures, both alone and in comparison with radiology residents.
Materials and methods
Consecutive recruitment of 366 children under 18 referred by the pediatric emergency department for radiographic exploration of trauma involving peripheral limbs (five subgroups: hand/wrist, elbow, shoulder, knee/leg, and foot/ankle). Two fracture-detection AI algorithms (Boneview and Rayvolve) were evaluated, with joint effusions and dislocations analyzed only by Boneview. Two radiology residents (first and third years) interpreted the radiographs first without and then with AI assistance.
Results
Boneview’s sensitivity and specificity were 0.85 and 0.87, respectively, while Rayvolve’s were 0.81 and 0.93. No significant difference was observed between areas under the curve of Boneview (0.862) and Rayvolve (0.874) with P-value = 0.62. AI significantly improved the performance of the least experienced resident, but not of the more experienced resident. For effusion detection, Boneview’s sensitivity and specificity were 0.57 and 0.96, close to but slightly lower than the performance of the residents. AI assistance did not have real impact on residents’ performance for effusion detection.
Conclusion
Evaluating two AI algorithms in the same pediatric population revealed consistent and acceptable performance with variations in performance profiles and may improve outcomes for less-experienced users.
