Multi-scale feature refinement network for lower limb fracture detection in X-ray images

Lower limb fractures are common orthopedic emergencies, and rapid and accurate diagnosis of such fractures is essential for clinical practice. However, challenges such as large variations in lesion scales and ambiguous boundaries between fracture foreground and background hinder precise detection of lower limb fractures.

To address these challenges, we propose a multi-scale feature refinement network, MFRNet. Two innovative modules are introduced: the Adaptive Feature Perception Block (AFPB) enhances fracture-related features and suppresses background noise via a bottleneck structure and a lightweight channel-spatial attention mechanism. The Multi-Scale Dilated Attention Module (MSDAM) expands the receptive field through parallel dilated convolutions and multi-head attention mechanism, thereby capturing rich multi-scale contextual information.

Extensive experimental results on the lower limb fracture X-ray dataset demonstrate that MFRNet achieves 89.1% mAP and 52.2% mAP50-95 with only 3.8 M parameters, outperforming current mainstream object detection models.

The proposed MFRNet strikes a favorable balance between high detection accuracy and parameter efficiency, offering clinical practical value for computer-aided diagnosis of lower limb fractures. Future work will focus on multi-center validation, fracture type classification, and mobile deployment.