【摘要】 目的 心包积液(pericardial effusion,PE)是指心包本身或者全身性疾病导致的心包内积液分泌过多，作为一种病理表征，目前临床中最常使用的心包积液检测方法是超声心动图技术。但该检测技术极大地依赖于操作者的临床经验，因此本研究开发一种用于定量测量超声心动图中PE的深度学习方法，以期实现超声心动图检查时心包积液测量的自动化。方法 首先利用残差网络(residual network, ResNet)进行4个探测视角左室短轴切面(left ventricular short-axis view, SAX)、右室流出道切面(right ventricular outflow tract view,RVOT)、胸骨旁左室长轴切面(parasternal long-axis view of the left ventricle,PLAX)、心尖四腔切面(apical four-chamber view,A4C)的分类；然后将VGG加入Unet网络中，利用迁移学习进行心包积液的自动分割，以戴斯相似系数、准确率、召回率对分割网络进行评价；最后，使用最大内切测量法从分割的标签中计算PE的最大深度。结果 图像分类模块中，4个视角的图像分类平均准确率为0.84;在超声心动图的心包积液自动分割模块中，测试集各视角的PE分割平均准确率为0.82;对所提积液测量方法的测量结果与医生手动标注结果进行一致性分析，最终结果得到标准测量值和机器测量值之间的类内相关系数(intraclass correlation coefficient,ICC)为0.928(95%CI:0.832～0.962),表现出较高的一致性。结论 本研究提出的方法在超声心动图自动分类、心包积液的自动分割和最大深度测量方面都表现出较好的性能，有望在未来的PE检测中提供一种标准化的客观方法，减少对医师的依赖性。更多还原

【Abstract】 Objective Pericardial effusion(PE) refers to the excessive secretion of fluid in the pericardium caused by pericardial or systemic diseases. As a pathological manifestation, echocardiography is currently the most commonly used method for detecting pericardial effusion in clinical practice. However, this detection technique highly depends on the operator’s clinical experience. Therefore, this study develops a deep learning method for quantitatively measuring PE in echocardiograms, aiming to achieve automation of pericardial effusion measurement during echocardiographic examination. Methods Firstly, the residual network(ResNet) is used to classify the four detection views: the left ventricular short-axis view(SAX),the right ventricular outflow tract view(RVOT),the parasternal long-axis view of the left ventricle(PLAX),and the apical four-chamber view(A4C). Then, VGG is incorporated into the Unet network, and transfer learning is utilized for the automatic segmentation of pericardial effusion. The segmentation network is evaluated using the Dice similarity coefficient, accuracy, and recall. Finally, the maximum inscribed measurement method is used to calculate the maximum depth of PE from the segmented labels. Results In the image classification module, the average accuracy for categorizing the four standard views is 0.84. For the automatic pericardial effusion(PE) segmentation module, the mean segmentation accuracy across test sets reach 0.82. Consistency analysis between the proposed automated measurement method and manual physician annotations yield an intraclass correlation coefficient(ICC) of 0.928(95% CI:0.832-0.962),demonstrating strong agreement. Conclusions This method exhibits robust performance in view classification, PE segmentation, and maximum depth quantification, offering a standardized and objective solution for future PE detection, thereby reducing reliance on manual physician assessments.更多还原