改进U-Net模型在心包积液超声心动图图像分割中的应用  

Application of improved U-Net model in the image segmentation of pericardial effusion echocardiography

作　　者：龚薇 李钟玉 苏振国 宁传永 铁晓烂 孙鹏 GONG Wei;LI Zhongyu;SU Zhenguo;NING Chuanyong;TIE Xiaolan;SUN Peng(School of Medical Information and Engineering,Ningxia Medical University,Yinchuan 750004,China)

机构地区：[1]宁夏医科大学医学信息与工程学院,银川市750004 [2]宁夏医科大学第一临床医学院 [3]宁夏回族自治区医学科学研究所

出　　处：《临床超声医学杂志》2024年第12期1051-1054,共4页Journal of Clinical Ultrasound in Medicine

基　　金：教育部产学合作协同育人项目(22107040035230);2023年宁夏回族自治区区级大学生创新创业训练计划项目(S202310752012)。

摘　　要：心包积液是由多种因素引起的心包腔内液体积聚,超声心动图是其常用的诊断工具,但由于图像噪声、回声变化和分割区域形状不规则等问题,常导致定位困难。本文提出了一种改进U-Net模型,并探讨其自动分割心包积液超声心动图图像的应用价值。本文纳入公开数据集Pericardial-Effusion-experimental-data中2541组(共5082张图片)数据,以7∶3的比例分为训练集(1779组,3558张图片)和公开测试集(762组,1524张图片),另纳入山东第二医科大学附属医院的心包积液患者超声心动图图像38张作为外部测试集。基于训练集对U-Net模型进行改进,具体方法包括引入多尺度特征提取模块和Dropout2d机制,以增强模型的泛化能力和分割精度;在下采样过程中应用LeakyReLU激活函数,提高模型的非线性表达能力;在卷积层中采用反射填充,优化积液区边界的轮廓。比较改进前后U-Net模型在公开测试集和外部测试集上的准确率、召回率、精确度和F1分数,结果显示,改进后U-Net模型在公开测试集和外部测试集中的准确率分别为96.97%和98.00%,召回率分别为91.47%和80.03%,精确度分别为69.84%和52.20%,F1分数分别为77.34%和60.86%。上述结果表明,改进U-Net模型具有较好的泛化能力,为心包积液超声心动图图像的自动分割提供了一种有效的解决方案,能够在提高诊断效率的同时,确保较高的准确率。Pericardial effusion refers to the accumulation of fluid in the pericardial cavity due to various causes.Although echocardiography is a common diagnostic tool,its effectiveness can be limited by issues such as image noise,echo variations,and irregular segmentation boundaries,making it challenging to accurately locate effusions.This study proposes an improved U-Net model to explore its application value in the automatic segmentation of pericardial effusion in echocardiographic images.The publicly available Pericardial-Effusion-experimental-data dataset,containing 2541 sets(5082 images),were divided into a 7∶3 ratio,with 1779 sets(3558 images)used for training and 762 sets(1524 images)for testing.Additionally,38 echocardiographic images of pericardial effusion patients from the Affiliated Hospital of Shandong Second Medical University were included as an external test set.The U-Net model was enhanced by incorporating a multi-scale feature extraction module and a Dropout2d mechanism to improve generalization and segmentation accuracy.The LeakyReLU activation function was applied during downsampling to boost the model’s nonlinear expression capability,while reflection padding was used in the convolutional layers to refine the boundaries of the effusion region.Performance comparison between the original and improved models,using the open and external test sets,revealed that the improved U-Net model achieved accuracy rates of 96.97%and 98.00%,recall rates of 91.47%and 80.03%,precision rates of 69.84%and 52.20%,and F1 scores of 77.34%and 60.86%,respectively.These results demonstrate that the improved U-Net model exhibits strong generalization ability and offers an effective solution for the automatic segmentation of pericardial effusion echocardiographic images,enhancing diagnostic efficiency while maintaining high accuracy.

关 键 词：超声心动描记术 心包积液 医学图像分割 U-Net模型 

分 类 号：R540.45[医药卫生—心血管疾病]