摘要
目的探讨基于U-Net神经网络的肥厚型心肌病(hypertrophic cardiomyopathy,HCM)与高血压性左心室肥厚(hypertensive left ventricular hypertrophy,HLVH)的磁共振图像定量分析与鉴别。材料与方法回顾性分析2017年国际医学图像计算和计算机辅助干预协会(Medical Image Computing and Computer Assisted Intervention Society,MICCAI)一项心脏疾病自动诊断挑战项目中包含的100例心脏疾病患者以及2014年7月至2019年3月上海交通大学医学院附属仁济医院确诊的45例HCM与48例HLVH患者。MICCAI数据集作为训练集和验证集,随机挑选5例HCM病例和5例HLVH病例作为测试集,得到一个基于U-Net的心脏自动分割神经网络。对所有入组的HCM与HLVH患者的心脏磁共振图像进行自动分割并提取多项量化参数,采用独立t检验比较各项量化参数在HCM组与HLVH组间的差异,采用多因素logistic回归法对有统计学差异的变量进一步分析建模,使用4折交叉验证方法结合ROC法对模型的分类性能进行验证。结果55项量化参数中有13项在HCM组与HLVH组之间存在显著性差异,有3项指标对两者的鉴别分类具有显著性影响。4折交叉验证得到的ROC曲线下面积分别为0.939、0.984、0.972和0.963,其中最佳模型对应的测试集准确率为86.96%(20/23)。结论U-Net神经网络分割心脏磁共振影像可以提供更多量化信息,有助于鉴别肥厚型心肌病与高血压性左心室肥厚。
Objective:To investigate the value of quantitative information of MRI got from U-Net neural network in the differentiation of hypertrophic cardiomyopathy and hypertensive left ventricular hypertrophy.Materials and Methods:We retrospectively analyzed 100 heart disease subjects collected from Medical Image Computing and Computer Assisted Intervention Society(MICCAI)2017 automated cardiac diagnosis challenge and 45 hypertrophic cardiomyopathy patients and 48 hypertensive left ventricular hypertrophy patients collected from July 2013 to March 2019 in the department of cardiology,Renji Hospital of Shanghai Jiaotong University.All patients underwent the steady state free precession cine sequence MRI scan in short axis.MICCAI dataset,separated into 1710 images and 190 images,were used as training dataset and validating dataset.Five hypertrophic cardiomyopathy patients and 5 hypertensive left ventricular hypertrophy patients,including 190 images,were selected as test dataset.The U-Net model was utilized in the segmentation of heart in cine MR images.The image segmentation was performed on all the hypertrophic cardiomyopathy and hypertensive left ventricular hypertrophy patients and the quantitative parameters were calculated based on the segmentation results.Independent t test was applied to compare the differences of all the parameters between the two diseases groups.Multivariate logistic regression and a 4-fold cross-validation method were applied to fit a diagnosis model and to validate the robust and diagnostic accuracy of the model.Results:Thirteen of all the 55 quantitative parameters had significant differences between the hypertrophic cardiomyopathy group and hypertensive left ventricular hypertrophy group,and 3 of them had significant influences on the classification between the two groups.The training set and the test set were 70 and 23 cases,and the areas under curves of ROC in test set produced from 4-fold cross-validation were 0.939,0.984,0.972 and 0.963.The accuracy of the test set corresponding to the best model was 86.96%(20/23).Conclusions:Automatic segmentation of heart in cine MR images based on U-Net neural network can provide more quantification information,which can help to diagnose the hypertrophic cardiomyopathy and hypertensive left ventricular hypertrophy.
作者
焦梓灵
魏寒宇
李继凡
陈硕
柴烨子
刘启明
李睿
姜萌
JIAO Ziling;WEI Hanyu;LI Jifan;CHEN Shuo;CHAI Yezi;LIU Qiming;LI Rui;JIANG Meng(Center for Biomedical Imaging Reseach,Department of Biomedical Engineering,School of Medicine,Tsinghua University,Beijing 100084,China;Department of Cardiology,Renji Hospital,School of Medicine,Shanghai Jiaotong University,Shanghai 200127,China)
出处
《磁共振成像》
CAS
2020年第9期741-746,共6页
Chinese Journal of Magnetic Resonance Imaging
基金
十三五国家重点研发计划(编号:2016YFC1301601)
十三五国家重点研发计划(编号:2017YFC0109002)
国家自然科学基金面上项目(编号:81971604)。
关键词
卷积神经网络
肥厚型心肌病
高血压性左心室肥厚
定量分析
磁共振成像
convolutional neural network
hypertrophic cardiomyopathy
hypertensive left ventricular hypertrophy
quantitative analysis
magnetic resonance imaging
