CRPA: You only look once with Cross-directional Regional Pooling Attention in mandibular fracture CT

Abstract Mandibular fracture detection is very important in clinical diagnosis. However, the CT images of mandibular fractures have the characteristics of complex fracture morphology, uneven location distribution and small fracture line, which makes it difficult for the YOLO model to fully model the structural characteristics and spatial dependence of the fracture area. In view of the above problems, this paper proposes an improved YOLO mandibular fracture detection model based on Pooling Attention Binary Relation (PABR). Firstly, a pooling attention binary relation representation framework is constructed, and the pooling response between different regions is uniformly represented as a weighted binary relation, and its properties are discussed. Secondly, based on this framework, a Cross-directional Regional Pooling Attention (CRPA) module is designed. By decoupling the horizontal and vertical regional pooling responses and coupling their cross-directional relationships, the proposed module enhances spatial dependencies while preserving positional information, thereby improving the representation capability of the model for complex fracture structures. Finally, CRPA is embedded in the YOLO network, and compared with the classical pooling attention module and the representative improvement methods in recent years. The experimental results show that the mAP50, mAP50-95, Recall, Precision and F1 values of the proposed method on the self-built mandibular fracture CT dataset reach 95.90%, 45.80%, 90.80%, 94.80% and 92.73%, respectively. All indicators are better than the baseline YOLO model and comparison method. This study is of great significance for the auxiliary diagnosis of mandibular fractures.