To propose a multi-scale feature refinement network, MFRNet, for improved detection of lower limb fractures in X-ray images.
Approach:
Adaptive Feature Perception Block (AFPB): Enhances fracture-related features and suppresses background noise using a bottleneck structure and lightweight channel-spatial attention mechanism.
Multi-Scale Dilated Attention Module (MSDAM): Expands the receptive field through parallel dilated convolutions and a multi-head attention mechanism to capture rich multi-scale contextual information.
Key Findings:
MFRNet achieves 89.1% mAP and 52.2% mAP50-95 with only 3.8 M parameters.
MFRNet outperforms current mainstream object detection models.
Interpretation:
MFRNet balances high detection accuracy with parameter efficiency.
Limitations:
Future work needed for multi-center validation.
Further research required for fracture type classification and mobile deployment.
Conclusion:
MFRNet significantly enhances the detection of lower limb fractures in X-ray images, addressing challenges faced by existing models.
