基于计算机视觉大模型的超声心动图左心室射血分数自动化测量

Automated Echocardiographic Measurement of Left Ventricular Ejection Fraction Based on Foundation Model in Computer Vision

目的:探索计算机视觉大模型在超声心动图中预测左心室射血分数(LVEF)的可行性。方法:基于目前最大的超声心动图视频公开数据集EchoNet-Dynamic,纳入10024例患者的10024段超声心动图动态视频。选用计算机视觉大模型VideoMAE V2,分别使用7460、1288、1276段视频对其进行微调、验证和测试。结果:VideoMAE V2预测LVEF与人类专家测量结果间的平均绝对误差为3.94%(95%CI:3.79%~4.11%),Pearson相关系数为0.91(95%CI:0.89~0.92)。此外,VideoMAE V2可准确判断患者LVEF是否低于50%,AUC可达0.96(95%CI:0.95~0.97)。结论:计算机视觉大模型用于在超声心动图视频中预测LVEF可行,为通用多模态超声心动图自动判读系统的开发奠定基础。

Objectives:To examine the feasibility of using foundation model in computer vision for echocardiographic left ventricular ejection fraction measurement.Methods:Based on the most extensive publicly accessible repository of echocardiographic loops,EchoNet-Dynamic,featuring 10024 recordings from individual patients,a foundation model in computer vision,VideoMAE V2,was fine-tuned,validated,tested using 7460,1288,and 1276 echocardiographic loops,respectively.Results:The mean absolute error between left ventricular ejection fraction measurements of VideoMAE V2 and expert's measurements was 3.94%(95%CI:3.79%-4.11%).The Pearson's correlation coefficient was 0.91(95%CI:0.89-0.92).Additionally,VideoMAE V2 demonstrated exceptional accuracy in identifying patients with a left ventricular ejection fraction below 50%,achieving an AUC of 0.96(95%CI:0.95-0.97).Conclusions:This study validates the feasibility of using foundation model in computer vision for measuring left ventricular ejection fraction in echocardiographic loops and lays the foundation for the development of a generalized multimodal automated interpretation system for echocardiography.