AI-assisted ultrasound for trauma: a scoping review.

Authors

Abstract

Artificial intelligence (AI) has become a disruptive technology in medicine, with rapidly developing implications for education, diagnosis, image enhancement and assessment, prognosis, and treatment. The literature on AI-assisted ultrasound is still developing. Therefore, we aimed to investigate how AI is employed in ultrasound for trauma and how AI-assisted ultrasound could be deployed in clinical practice. The review was conducted according to JBI's scoping review methodology. An appropriate search strategy was initially formulated for PubMed, Scopus, Web of Science (WoS), and Google Scholar (backward and forward citation tracking). The concepts of "artificial intelligence", "ultrasound", and "trauma" were searched across these databases using their respective Medical Subject Headings (MeSH). The core concepts were AI applications integrated with ultrasound in trauma. In this context, studies focusing on the following categories relating to the use of AI-assisted ultrasound in the field of trauma have been included in the review: (1) diagnostic accuracy, (2) differential diagnosis and confirmation of diagnosis, (3) quality and standardization, (4) operator support, and (5) training and modelling. Data extraction and analysis were performed using Microsoft Excel and were verified. Twenty-one studies were included. Targeted AI tasks primarily comprised diagnostic abnormality detection and segmentation/localization. Predominantly focused assessment with sonography for trauma (FAST), thoracic trauma evaluation, and extremity trauma. Although high apparent diagnostic accuracy was reported across development-stage retrospective cohorts, independent external validation was executed in only one study, with the remainder bounded by single-center data or synthetic phantoms. Crucially, none of the studies assessed subsequent patient outcomes such as mortality, time to intervention, or resource utilization. This scoping review confirms that while deep learning frameworks show promising retrospective diagnostic performance and image-quality automation, their translation into acute trauma care faces critical technical and clinical bottlenecks. Key current limitations include restricted model interpretability, missing regulatory pathways, a lack of clinician trust, inherent decision-making biases, poor algorithmic transparency, and the lack of patient-outcome tracking. To address these barriers, future studies should be designed with stronger methodological rigor, including prospective, multicenter, randomized controlled, and externally validated approaches. No clinical trial number is available.

Tags

Topics