PEYOLO: a wrist fracture detection network based on multi-level receptive field feature extraction and cross-scale fusion

Introduction: The wrist is a high-incidence site for traumatic injuries in the skeletal system, where rapid and accurate diagnosis of fractures is critically important in clinical practice. However, the complex multi-scale variations of fractures impede precise detection of wrist fractures. Methods: To address this issue, this study proposes a novel wrist fracture detection model, PEYOLO, based on multi-level receptive field feature extraction and cross-scale feature fusion. First, we design a Parallel Dilated Multi-head Attention Module (PDMAM), which performs sparse sampling under different receptive fields to simultaneously obtain receptive fields of varying scales within the same layer. Second, we introduce an Efficient Multi-Scale Attention module, which enhances the perception of multi-scale fracture features by fusing multi-scale spatial information with cross-dimensional dependencies. Results: Experimental results on the GRAZPEDWRI-DX-fracture dataset demonstrate that PEYOLO improves mAP by 1.4% over the baseline and outperforms several state-of-the-art object detection models. Discussion: PEYOLO has demonstrated highly advantageous high precision and inference speed, making it an important tool for clinical diagnosis and treatment planning.