Background: Rapid and accurate interpretation of trauma imaging is essential for timely clinical decision-making, yet high workload and complex injury patterns can increase diagnostic challenges. This systematic review aimed to synthesize current evidence on deep learning (DL) applications in trauma radiology, with particular emphasis on diagnostic performance, generalizability, workflow integration, and clinical translation.
Methods: In accordance with PRISMA 2020, PubMed, Scopus, Web of Science, and Embase were systematically searched for eligible studies evaluating DL-based interpretation of radiological images in traumatic injuries. Data on injury type, imaging modality, DL architecture, study design, validation strategy, diagnostic performance, workflow outcomes, interpretability, and clinical integration were extracted independently by two reviewers. Methodological quality was assessed using QUADAS-2. Because substantial heterogeneity existed in populations, injury types, imaging modalities, reference standards, architectures, and validation strategies, quantitative meta-analysis and network meta-analysis were not performed. Instead, evidence was synthesized using a structured qualitative approach and used to develop a conceptual clinical-translation framework, clinical-readiness assessment, and phased implementation roadmap.
Results: Forty-eight studies met the eligibility criteria. DL models, particularly CNN-based architectures, demonstrated generally high diagnostic performance for intracranial hemorrhage, fractures, and chest trauma, with reported AUCs frequently exceeding 0.90. AI-assisted workflows were associated with reductions in interpretation and notification times. However, performance frequently declined during external validation, with reported AUC reductions of approximately 5–15%. Risk of bias and limited reporting of prospective, multicenter, and real-world evaluation remained important concerns. Clinical translation was further constrained by dataset shift, limited interpretability, workflow integration challenges, and insufficient evidence regarding regulatory and economic implementation.
Conclusion: DL shows considerable promise as an adjunct to trauma radiology, but high reported diagnostic accuracy alone is insufficient to establish clinical readiness. Future research should prioritize multicenter external validation, explainable AI, prospective workflow evaluation, and continuous post-deployment monitoring to support safe and generalizable clinical implementation.