Enhancing Bone Fracture Detection with Large Kernel Attention Modules

Accurately diagnosing complex fractures in hospital emergency departments remains a significant challenge, as it requires specialized and rapid decision-making in a fast-paced environment. While automated fracture detection systems have gained traction thanks to advancements in computer vision, few implement domain-inspired insights to guide architectural design. In this paper, we introduce a novel Large Kernel Attention Module (LKAM), inspired by the diagnostic behavior of radiologists who often use changes related to adjacent tissues to localize hidden fractures. Unlike existing attention mechanisms, LKAM employs large convolutional kernels combined with a channel and spatial attention mechanism to expand the receptive field significantly, enabling the model to capture valuable information beyond the immediate fracture location. The LKAM was integrated into the You Only Look Once v5 (YOLOv5) architecture, using different backbones, and its performance was validated through experiments on the FractAtlas dataset. The present study outperformed other state-of-the-art attention modules, achieving a mAP (0.5) of 58.0 with CSPDarknet53, and 56.1 with EfficientNet as the backbone.