PSCA-YOLO: An enhanced YOLO with position-semantics coupled attention for mandibular fracture detection

Mandibular fractures are a common type of maxillofacial trauma, and accurate localization and classification are crucial for clinical diagnosis. However, the complex morphology of fractures limits existing detection models in effectively representing semantic and positional features. To address this problem, this study proposes PSCA-YOLO, which integrates feature decoupling and coupling mechanisms into YOLO. First, a localization-classification dual-branch decoupling structure is designed, where the classification branch is supervised by binary cross-entropy loss (BCE) to enhance semantic representation, while the localization branch combines distributed focal loss (DFL) and CIOU loss to strengthen spatial learning. Second, a position-semantics feature coupling attention module is introduced to fuse semantic and positional features, improving feature perception. Experiments on mandibular fracture CT dataset, demonstrate strong effectiveness. PSCA-YOLO improves mAP50 by 2.60%, recall by 2.20%, and precision by 2.90% compared with the baseline model. Therefore, the model has positive significance for the clinical diagnosis of mandibular fractures.