基于改进YOLOv5m的血管介入导丝检测算法
Vascular Intervention Guidewire Detection Algorithm Based on Improved YOLOv5m
鲍芳嘉 1张明伟 1张天逸 1程云章1
作者信息
- 1. 上海理工大学 健康科学与工程学院;上海介入医疗器械工程技术研究中心,上海 200093
- 折叠
摘要
针对目前主流目标检测算法在微创血管介入导丝检测中应用较少、检测准确率低和检测速度慢等问题,提出了一种改进的YOLOv5m网络用于血管介入导丝检测.首先,在YOLOv5m的主干网络中引入可变形卷积替换部分标准卷积,并在主干网络的CSP模块中添加坐标注意力机制;其次,在颈部采用加权双向特征金字塔网络进行特征融合,提高模型对不同特征层的融合能力.实验结果表明,改进YOLOv5m算法的mAP@0.5达到87.8%,比YOLOv5m提升了5.7%,表明该算法在血管介入导丝检测方面具有较高应用价值.
Abstract
In order to solve the problems of current mainstream target detection algorithms being less used in minimally invasive vascular in-terventional guidewire detection,low detection accuracy and slow detection speed,an improved YOLOv5m network is proposed for the detec-tion of vascular interventional guidewires.First,deformable convolution is introduced into the backbone network of YOLOv5m to replace some standard convolutions,and a coordinate attention mechanism is added to the CSP module of the backbone network;BiFPN is used in the neck to performs feature fusion improving the model's ability to fuse different feature layers.Experimental results show that the mAP@0.5 of the im-proved YOLOv5m algorithm reaches 87.8%,which is 5.7%higher than YOLOv5m,indicating that this algorithm has relatively high applica-tion value in vascular interventional guidewire detection.
关键词
血管介入/YOLOv5m/导丝检测/可变形卷积/坐标注意力机制/BiFPNKey words
vascular intervention/YOLOv5m/guidewire detection/deformable convolution/coordinate attention mechanism/BiFPN引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据