Artificial intelligence (AI) in radiological paediatric fracture assessment: an updated systematic review.

Background: Recognising bone injuries in children is a critical part of children's imaging, and, recently, several AI algorithms have been developed for this purpose, both in research and commercial settings. We present an updated systematic review of the literature, including the latest developments.

Methods/materials: Scopus, Web of Science, Pubmed, Embase, and Cochrane Library databases were queried for studies published between 1 January 2011 and 6 September 2024 matching search terms 'child', 'AI', 'fracture,' and 'imaging'. Retrieved studies were evaluated, and descriptive statistics were collated for diagnostic performance.

Results: Twenty-six eligible articles were included; seventeen (17/26, 65.%) of these were published within the last two years. Six studies (6/26, 23.1%) used open-source datasets to train their algorithm, the remainder used local data. Sixteen studies (16/26, 61.5%) evaluated a single joint (wrist, elbow, or ankle); multiple bones within the appendicular skeleton were assessed in the other ten studies. Seven articles (7/26, 26.9%) related to the performance of a commercial AI tool. Accuracy of AI models ranged from 85.0 to 100.0%. Six studies (6/26, 23.1%) evaluated the accuracy of human readers with and without AI assistance, of which two studies found a statistically significant improvement when humans were assisted by AI. The largest pool of human readers in any paper consisted of 11 readers of varying experience.

Conclusion: The pace of research in AI fracture detection in children's imaging has increased. Studies show high accuracy of AI models, but proof of clinical impact, cost-effectiveness, and any socioeconomic or ethical bias are still lacking.

Key Points: Question AI model development has rapidly increased in recent years. We present the latest developments in AI model diagnostic accuracy for paediatric fracture detection. Findings Studies now demonstrate performance improvement when AI is used to assist human interpretation of paediatric fractures, especially when aiding junior radiologists. Clinical relevance Studies show high accuracy for AI models; however, further research is needed to evaluate AI across diverse age groups, bone diseases, and fracture types. Evidence of real-world patient benefit for AI and any socioeconomic or ethical bias are still lacking.