摘要
目的基于U形全卷积神经网络(U-Net)深度学习模型,探讨急诊CT图像中自发性脑实质出血病灶自动分割及体积测量的可行性。资料与方法回顾性收集2009年7月25日—2019年11月6日北京大学第一医院急诊头颅CT检查诊断为脑实质出血的CT影像和报告,共256例数据,由3名影像医师标注颅内血肿病灶。将256例数据随机分为训练集206例、调优集25例和内部验证集25例训练3D U-Net分割模型。另收集2019年11月7日—2020年3月31日共50例数据用于外部验证。以内、外部验证集中血肿病灶的Dice相似系数(DSC)评价模型的分割效果,并自动生成血肿的三维径线和体积定量值。比较模型预测的定量值(模型值)、既往临床报告中的定量值(报告值)和专家标注的定量值(参考值),评价不同测量方法所得定量值之间的一致性。结果在内部验证集25例中U-Net模型检出全部血肿,模型预测的敏感度为100%,以病灶为单位平均DSC为0.84;在外部验证集50例中,U-Net模型检出49例,以病灶为单位平均DSC为0.90。在内部验证集中模型值与参考值比较,血肿病灶三维径线和体积差异均无统计学意义(P>0.05)。报告值与参考值比较,左右径及前后径差异有统计学意义(Z=-4.319、-3.242,P<0.05),上下径和体积差异无统计学意义(P>0.05)。在外部验证集中模型值与参考值比较,血肿病灶的上下径及体积差异有统计学意义(Z=-2.146、-2.590,P<0.05),左右径及前后径差异均无统计学意义(P>0.05)。报告值与参考值比较,血肿病灶三维径线和体积差异均有统计学意义(Z=-4.793、-4.580、-5.855、-3.335,P<0.05)。内、外部验证集中,模型值和报告值与参考值的一致性均高,两两比较组内相关系数均达到0.90以上。结论深度学习模型可用于急诊CT图像中自发性脑实质出血的自动分割及体积测量。
Purpose To explore the feasibility of automatic segmentation and volume measurement of intracranial hemorrhage on emergency CT images based on the U-shaped fully convolutional neural network,U-Net deep learning model.Materials and Methods The CT images and reports of acute intracranial hemorrhage were collected retrospectively from July 25,2009 to November 6,2019 in Peking University First Hospital emergency head CT scanning.A total of 256 cases were included in this study.Three radiologists marked intracranial hematoma lesions.A 3D U-Net model was trained to segment the hematomas,with randomly divided into training set(n=206),fine tuning set(n=25),and internal validation set(n=25).Another dataset including 50 cases was collected from November 7,2019 to March 31,2020 for the external validation.Dice similarity coefficient(DSC)values of the hematomas lesion in test set of the internal and external validation were all used to evaluate the segmentation effect of the model,and the three-dimensional diameter and volume quantitative value of the hematoma were automatically generated.The quantitative value predicted by the model(model value),the quantitative value reported in the previous clinical practice(reported value),and the quantitative value marked by experts(ground truth)were compared,and the consistency of the quantitative values obtained by different measurement methods was tested.Results In the internal validation group,all the hematomas(n=25)were fully detected by U-Net model,and the sensitivity of the model was 100%,and the average DSC was 0.84.Of the 50 cases,a total of 49cases with hematomas were detected in the external validation group,and the average DSC was 0.90.There were no statistically significant differences in the three-dimensional diameter and volumes of hematoma lesions between the model value and ground truth in the internal validation group(P>0.05).Compared reported value with the ground truth,the differences between left and right diameters and front and rear diameters were statistically si
作者
奈日乐
王可欣
谢辉辉
杨洁瑾
蔡金秀
李昌欣
王祥鹏
张晓东
王霄英
NAI Rile;WANG Kexin;XIE Huihui;YANG Jiejin;CAI Jinxiu;LI Changxin;WANG Xiangpeng;ZHANG Xiaodong;WANG Xiaoying(Department of Radiology,Peking University First Hospital,Beijing 100034,China;不详)
出处
《中国医学影像学杂志》
CSCD
北大核心
2022年第11期1089-1094,1101,共7页
Chinese Journal of Medical Imaging
关键词
颅内血肿
体层摄影术
X线计算机
深度学习
分割
结构化报告
Intracranial hemorrhage
Tomography
X-ray computed
Deep learning
Segmentation
Structured report
