基于U-Net神经网络的多模态MR颈动脉血管成像的分割方法研究  

The study on the segmentation of carotid vessel wall in multicontrast MR images based on U-Net neural network

作　　者：李继凡 陈硕[1] 章强[1] 宋焱[2] Gador Canton 孙杰[3] 许东翔 赵锡海[1] 苑纯[1] 李睿[1] Li Jifan;Chen Shuo;Zhang Qiang;Song Yan;Gador Canton;Sun Jie;Xu Dongxiang;Zhao Xihai;Yuan Chun;Li Rui(Center for Biomedical Imaging Research,Department of Biomedical Engineering,School of Medicine,Tsinghua University,Beijing 100084,China;Department of Radiology,Beijing Hospital,Beijing 100730,China;Vascular Imaging Laboratory,Department of Radiology,University of Washington,Seattle,WA 90876,United States)

机构地区：[1]清华大学医学院生物医学工程系生物医学影像研究中心,北京100084 [2]北京医院放射科,100730 [3]美国华盛顿大学放射学系血管成像实验室,西雅图90876

出　　处：《中华放射学杂志》2019年第12期1091-1095,共5页Chinese Journal of Radiology

基　　金：国家重点研发计划(2016YFC1301601)。

摘　　要：目的 探讨基于U-Net神经网络的多模态MR影像颈动脉血管分割方法的价值.方法 回顾性分析了2012年至2015年中国动脉粥样硬化风险评估研究项目中,经标准多模态MR扫描,且两周内出现缺血性脑卒中或短暂性脑缺血的患者.经纳入标准和排除标准筛选后,有658例患者共17 568层颈动脉血管壁影像纳入研究.应用定制设计的心血管疾病评估计算机辅助系统(CASCADE,华盛顿大学血管成像实验室,西雅图)对所有影像数据进行分析.按照训练集、验证集和测试集6∶2∶2的比例,随机选取10 592个样本作为训练集,3 488个样本作为验证集,3 488个样本作为测试集.为防止模型过拟合,提高模型泛化能力,对原始的多模态血管斑块MR影像进行数据增强.应用经过微调的U-Net神经网络构建多模态MR影像颈动脉血管分割模型,在训练集上训练,在验证集上验证并优化训练超参数,在测试集上测试并计算像素级别的颈动脉血管分割的敏感度、特异度和Dice系数,并计算U-Net分割方法和手工分割方法下的最大管壁厚度和管壁面积,利用组内相关系数和Bland-Altman分析来验证两种方法的一致性.结果 在测试集上应用训练得到的U-Net神经网络模型进行颈动脉血管分割,计算敏感度为0.878,特异度为0.986,Dice系数为0.858.最大管壁厚度的组内相关系数(95%可信区间)为0.921(0.915~0.925),管壁面积的组内相关系数(95%可信区间)为0.929 (0.924~0.933),Bland-Altman分析中最大管壁厚度差值为(0.037±0.316)mm,管壁面积差值为(1.182± 4.953)mm2,U-Net分割方法和手工分割方法具有较高一致性.结论 应用U-Net神经网络的方法,在大规模经过专业医师标注的数据集上进行训练和验证,可以实现对多模态MR影像颈动脉血管自动分割.Objective To investigate the value of automatic segmentation of carotid vessel wall in multicontrast MR images using U-Net neural network.Methods Patients were retrospectively collected from 2012 to 2015 in Carotid Atherosclerosis Risk Assessment(CARE II)study.All patients who recently suffered ischemic stroke and/or transient ischemic attack underwent identical,state-of-the-art multicontrast MRI technique.A total of 17568 carotid vessel wall MR images from 658 subjects were included in this study after inclusion criteria and exclusion criteria.All MR images were analyzed using customized analysis platform(CASCADE).Randomly,10592 images were assigned into training dataset,3488 images were assigned into validating dataset and 3488 images were assigned into test dataset according to a ratio of 6∶2∶2.Data augmentation was performed to avoid over fitting and improve the ability of model generalization.The fine-tuned U-Net model was utilized in the segmentation of carotid vessel wall in multicontrast MR images.The U-Net model was trained in the training dataset and validated in the validating dataset.To evaluate the accuracy of carotid vessel wall segmentation,the sensitivity,specificity and Dice coefficient were used in the testing dataset.In addition,the interclass correlation and the Bland-Altman analysis of max wall thickness and wall area were obtained to demonstrate the agreement of the U-Net segmentation and the manual segmentation.Results The sensitivity,specificity and Dice coefficient of the fine-tuned U-Net model achieved 0.878,0.986 and 0.858 in the test dataset,respectively.The interclass correlation(95%confidence interval)was 0.921(0.915-0.925)for max wall thickness and 0.929(0.924-0.933)for wall area.In the Bland-Altman analysis,the difference of max wall thickness was(0.037±0.316)mm and the difference of wall area was(1.182±4.953)mm2.The substantial agreement was observed between U-Net segmentation method and manual segmentation method.Conclusion Automatic segmentation of carotid vessel wall in m

关 键 词：神经网络 颈动脉 动脉粥样硬化 磁共振成像 

分 类 号：R74[医药卫生—神经病学与精神病学] R44[医药卫生—临床医学]