Diagnostic performance of artificial intelligence for facial fracture detection: a systematic review

To evaluate the diagnostic performance of artificial intelligence (AI) models for detecting facial bone fractures on computed tomography (CT), cone-beam CT (CBCT), and plain radiographs.

Original studies applying machine learning or deep learning algorithms for facial fracture detection in humans were included if they reported diagnostic accuracy metrics such as sensitivity, specificity, or area under the curve (AUC). PubMed-MEDLINE, Scopus, and Web of Science databases were searched up to June 3, 2025. Risk of bias was assessed using the QUADAS-2 tool. The review followed PRISMA 2020 guidelines and was registered in PROSPERO (CRD420251085644).

A total of 23 studies were included. Object detection models such as YOLOv5 and Faster R-CNN-demonstrated high diagnostic accuracy in localizing facial fractures. Classification models such as ResNet and Swin Transformer achieved AUCs frequently exceeding 0.90. Segmentation and hybrid frameworks further improved anatomical specificity. However, the generalizability of findings was constrained by predominantly retrospective, single-centre study designs, limited sample sizes, inconsistent annotation practices, and the absence of external or prospective validation.

AI models show high diagnostic performance for detecting facial fractures across multiple anatomical regions and imaging modalities. Further multicentre prospective studies and the integration of explainable AI are essential for clinical adoption.

AI-assisted diagnostic models have the potential to enhance facial fracture detection accuracy, especially in emergency and resource-limited settings. Their integration into radiology workflows could reduce interpretation time, support less experienced clinicians, and improve patient outcomes.