Detection of femoral neck fractures based on the PCSD-YOLO model

Femoral neck fractures are associated with high morbidity and mortality, and accurate early detection on radiographs is critical for clinical decision-making. However, existing automated detection models often suffer from limited accuracy, missed diagnoses, and insufficient sensitivity to small or irregular fracture patterns. To address these challenges, we propose PCSD-YOLO, an improved detection framework built upon YOLOv11n. The proposed model introduces three key modifications: (1) a Position–Channel Attention and Channel Aggregation (PACA) mechanism embedded within a C2PACA module to strengthen multi-scale channel representation and gated local–global positional modeling; (2) deformable convolution was introduced to replace the fixed depthwise-convolution-based positional encoding operation in the original C2PSA attention block; and (3) integration of the SPD-Conv module in the backbone to preserve fine-grained multi-scale information while reducing computational overhead. Experiments conducted on a retrospectively collected multi-hospital femoral neck fracture X-ray dataset demonstrate that PCSD-YOLO achieves a precision of 85.1 %, recall of 80.5 %, [email protected] of 85.0 %, and an F1-score of 82.7 %, with a 9.68 % reduction in model parameters compared to the baseline. Comparative evaluations show consistent improvements in both detection accuracy and localization performance over existing state-of-the-art models. These results indicate that PCSD-YOLO provides an effective and computationally efficient solution for automated femoral neck fracture detection, with potential value in assisting early clinical diagnosis and treatment planning.