基于改进YOLOv8的超声心动图像中心脏瓣膜检测算法  

The detection algorithm of cardiac valves in echocardiography based on improved YOLOv8

作　　者：吕洁 王文娟 王军 LYU Jie;WANG Wenjuan;WANG Jun(Taiyuan Central Hospital,Taiyuan 030009;Shanxi International Travel Health Care Center,Taiyuan 030021)

机构地区：[1]太原市中心医院/山西医科大学第九临床医学院,太原030009 [2]山西国际旅行卫生保健中心,太原030021

出　　处：《北京生物医学工程》2025年第2期149-157,共9页Beijing Biomedical Engineering

基　　金：山西省卫生健康委科研项目(2020109)资助。

摘　　要：目的在超声心动图像检测中,心脏瓣膜检测面临复杂背景干扰、目标小等问题,容易出现漏检、误检等。本文提出了一种用于超声心动图动态检测中心脏瓣膜实时识别与定位的改进YOLOv8算法。方法在改进YOLOv8算法中,基于YOLOv8模型进行优化,引入深度分离卷积扩大感受野,有效捕捉不同尺度的特征并保留细粒度空间信息,提升特征传递能力,降低误检率;采用坐标注意力模块(coordinate attention,CA)将位置信息整合到通道注意力中,使网络能更有效地关注广泛区域,增强细粒度特征提取,提高检测准确性和泛化性能;通过多尺度特征增强模块(multi-scale feature enhancement module,MSFE)利用多分支卷积和多尺度连接多通道特征图以增加网络对小尺寸目标的适应性;使用α-Focal EIoU损失函数提高边界框回归精度,增强模型对小目标的关注。结果针对以上模块的消融实验中,各模块的引入和优化显著提升了算法整体性能。改进后的YOLOv8模型在三尖瓣、二尖瓣和主动脉瓣检测中,mAP相比于基线YOLOv8提高6.9%,达到98.1%,其中,精确度为98.6%,召回率为97.6%,推理速度为33.2 FPS。结论相比于YOLOv5、SSD、Faster RCNN、RCNN和DETR等算法,改进YOLOv8算法在性能上有明显提升。基于消融实验分析可知,优化损失函数、坐标注意力模块和多尺度特征增强模块显著提升了模型性能,深度分离卷积网络则主要提升了计算效率。改进YOLOv8算法有助于超声图心脏瓣膜快速检测的自动化实现,为后续基于心脏瓣膜的医学分析奠定基础。Objective In echocardiographic image detection,heart valve detection faces problems such as complex background interference and small targets,which are prone to missed detections,false detections and other problems.This paper proposes an improved YOLOv8 algorithm for real-time identification and positioning of heart valves in dynamic echocardiography detection.Methods In the improved YOLOv8 algorithm,optimization is based on the YOLOv8 model,depth separable convolution is introduced to expand the receptive field,effectively capture features of different scales and retain fine-grained spatial information,improve feature transfer capabilities,and reduce false detection rates.The coordinate attention(CA)is used to integrate position information into channel attention,so that the network can more effectively focus on a wide area,enhance fine-grained feature extraction,and improve detection accuracy and generalization performance.The multi-scale feature enhancement module(MSFE)uses multi-branch convolution and multi-scale connection of multi-channel feature maps to increase the network's adaptability to small-sized targets.Theα-Focal EIoU loss function is used to improve the bounding box regression accuracy and enhance the model's focus on small targets.Results In the ablation experiments of the above modules,the introduction and optimization of each module significantly improves the overall performance of the algorithm.In the detection of tricuspid valve,mitral valve and aortic valve,the improved YOLOv8 model's mAP increases by 6.9%compared to the baseline YOLOv8,reaching 98.1%,of which Precision is 98.6%,Recall is 97.6%,and the inference speed is 33.2 FPS.Conclusions Compared with algorithms such as YOLOv5,SSD,Faster RCNN,RCNN and DETR,the improved YOLOv8 algorithm significantly improves performance.Based on ablation experimental analysis,it can be seen that the optimized loss function,coordinate attention module and multi-scale feature enhancement module are significantly improved.The model performance is improved

关 键 词：超声心动图 心脏瓣膜检测 YOLOv8 坐标注意力 多尺度特征增强 

分 类 号：R318.04[医药卫生—生物医学工程]