Enhancing Bone Fracture Detection with Large Kernel Attention Modules.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Diagnosing complex fractures in emergency departments is challenging due to the need for rapid, specialized decision-making.
• Existing automated fracture detection systems lack domain-inspired architectural insights.
• Radiologists often identify fractures by analyzing adjacent tissue changes.

Purpose of the Study:

• To introduce a novel Large Kernel Attention Module (LKAM) for enhanced automated fracture detection.
• To integrate LKAM into the YOLOv5 architecture, inspired by radiologist diagnostic behavior.
• To improve the accuracy and efficiency of fracture detection in medical images.

Main Methods:

• Developed a Large Kernel Attention Module (LKAM) using large convolutional kernels and channel/spatial attention.
• Integrated LKAM into the YOLOv5 object detection architecture with CSPDarknet53 and EfficientNet backbones.
• Validated the model's performance on the FractAtlas dataset.

Main Results:

• The LKAM significantly expanded the model's receptive field, capturing information beyond immediate fracture locations.
• The integrated YOLOv5-LKAM model outperformed other state-of-the-art attention modules.
• Achieved a mean Average Precision (mAP) of 58.0 with CSPDarknet53 and 56.1 with EfficientNet.

Conclusions:

• The LKAM is an effective attention mechanism for improving automated fracture detection systems.
• Domain-inspired insights can significantly enhance the architectural design of computer vision models for medical applications.
• This novel approach shows promise for more accurate and efficient fracture diagnosis in clinical settings.