EchoGPK:基于先验知识引导的超声心动图轻量级图卷积分析方法

EchoGPK:A Lightweight Graph Convolutional Analysis Method for Echocardiography Based on Prior Knowledge Guidance

根据超声心动图准确分析左心室轮廓和射血分数对于心血管疾病诊断意义重大.但现有方法存在左心室分割和射血分数预测之间缺乏关联性、左心室分割关键点易于出现离群点和突变点、方法存储和计算开销大、解释性不佳等问题,为此提出一种基于先验知识引导的轻量级图卷积方法EchoGPK(Echo Guided by Priori Knowledge),以心脏的结构和运动特性、相邻心肌的相似性等先验知识为引导,设计了计算高效的螺旋聚合函数和深度压缩的多头偏心聚合解码器,实现了图卷积结构的轻量化.方法基于临床医生的普遍经验提出了适度利用左心室轮廓的多任务射血分数预测网络,建立了左心室分割和射血分数预测之间的关联性,增强了推理的可解释性;基于图卷积神经网络的传递特性约束邻居点的行为,减少了边界离群点和突变点的产生.EchoGPK在大型公开数据集EchoNet-Dynamic上的实验结果表明,左心室分割的Dice分数达92.13%,射血分数预测的MAE达3.92%;方法表现出准确率高、参数量和算力需求低等特点,证明了先验知识在超声医学图像分析中的有效性.

Accurate analysis of the left ventricular outline and ejection fraction through echocardiography holds sig-nificant diagnostic implications in cardiovascular diseases.However,current methodologies exhibit deficiencies such as a lack of correlation between left ventricular segmentation and ejection fraction prediction,susceptibility to outliers and abrupt variations in key points of left ventricular segmentation,substantial storage and computational overhead,and poor in-terpretability.In addressing these issues,this study proposes a lightweight graph convolutional method termed EchoGPK(Echo Guided by Priori Knowledge).Guided by prior knowledge encompassing cardiac structure,motion characteristics,and the similarity among adjacent myocardial regions,the approach incorporates a computationally efficient spiral aggrega-tion function and a deeply compressed multi-head eccentric aggregation decoder,achieving the lightweighting of the graph convolutional structure.Leveraging the common experiences of clinical practitioners,the method introduces a multi task ejection fraction prediction network that moderately utilizes left ventricular contours,establishing a correlation between left ventricular segmentation and ejection fraction prediction to enhance interpretability.By employing the graph convolutional neural network transmission characteristics to constrain the behavior of neighboring points,the generation of boundary outli-ers and abrupt variations is reduced.Experimental results on the large-scale public dataset EchoNet-Dynamic demonstrate that EchoGPK achieves a Dice score of 92.13%for left ventricular segmentation and a mean absolute error(MAE)of 3.92%for ejection fraction prediction.Furthermore,the method exhibits higher accuracy,superior parameter count and computa-tional efficiency compared to relevant approaches,affirming the effectiveness of prior knowledge in ultrasound medical im-age analysis.