PSCA-YOLO: An enhanced YOLO with position-semantics coupled attention for mandibular fracture detection

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Mandibular fractures are frequent maxillofacial traumas requiring precise diagnosis.
• Complex fracture morphology challenges current detection models' ability to represent semantic and positional features.
• Accurate localization and classification are vital for effective clinical management.

Purpose of the Study:

• To develop an advanced deep learning model for improved mandibular fracture detection.
• To enhance the representation of both semantic and positional features in fracture detection.
• To improve the accuracy and efficiency of diagnosing mandibular fractures.

Main Methods:

• Proposed PSCA-YOLO model integrating feature decoupling and coupling mechanisms into YOLO.
• Designed a localization-classification dual-branch decoupling structure with specific loss functions (BCE, DFL, CIOU loss).
• Introduced a position-semantics feature coupling attention module for enhanced feature fusion.

Main Results:

• PSCA-YOLO demonstrated strong effectiveness on a mandibular fracture CT dataset.
• Achieved a 2.60% improvement in mAP50, 2.20% in recall, and 2.90% in precision compared to the baseline model.
• The model successfully enhanced semantic representation and spatial learning for fracture detection.

Conclusions:

• PSCA-YOLO significantly improves the detection performance for mandibular fractures.
• The proposed model offers positive implications for clinical diagnosis of maxillofacial trauma.
• Feature decoupling and coupling mechanisms are effective in addressing complex fracture morphology challenges.