利用U-net算法在头CT平扫图像上分割脑梗死的初步探究  

A preliminary study of cerebral infarction segmentation on CT images based on U-Net algorithm

作　　者：李晓庆 王可欣 额·图娅 李昌欣 王祥鹏 张晓东[1] 王霄英[1] LI Xiao-qing;WANG Ke-xin;E Tu-ya(Department of Radiology,Peking University First Hospital,Beijing 100034,China)

机构地区：[1]北京大学第一医院医学影像科,北京100034 [2]首都医科大学基础医学院,北京100069 [3]北京赛迈特锐医疗科技有限公司,北京100011

出　　处：《放射学实践》2022年第6期669-675,共7页Radiologic Practice

摘　　要：目的:基于深度学习方法训练辅助诊断模型,探究对头CT平扫(NCHCT)图像中脑梗死病灶自动分割的可行性。方法:搜集自2018年5月1日-2020年8月31日影像诊断报告中诊断印象包含“梗死”关键词的NCHCT连续病例1185例,筛选出最终证实为脑梗死的数据362例。由两位医师结合最终结果手工标注脑梗死区域。将数据按8:1:1的比例随机分为训练集(training set,n=288)、调优集(validate set,n=37)和测试集(test set,n=37例)。训练2D U-net模型分割脑梗死病灶,分割结果自动输入到结构化报告中。对测试集的预测结果和人工标注的结果进行比较,统计梗死病灶体积、径线的差异,使用Dice相似系数(DSC)、体积相似度(VS)和Hausdorff距离(HD)评价模型的预测效能。使用Bland-Altman评价模型预测的病灶体积、径线和CT值与手工标注的一致性。结果:测试集中平均DSC为0.66(95%CI:0.60~0.72),平均VS为0.75(95%CI:0.69~0.82),平均HD为39.69 mm(95%CI:32.38~47.01)。Bland-Altman图显示模型预测与手工标注对病灶大小和CT值测量的一致性较高,体积、径线和CT值数据点位于95%一致性界限(95%limits of agreement,95%LoA)外的数据为2.8%~11.1%。结论:基于深度学习的辅助诊断模型可用于分割NCHCT中的脑梗死病灶,并自动生成报告,对患者分诊有一定作用。Objective:To explore the feasibility of cerebral infarction segmentation on non-contrast head CT(NCHCT)images by using deep learning algorithms.Methods:Totally 1185 NCHCT cases,diagnosed as"infarction"in radiologic reports,were retrospectively collected from May 1,2018,to August 31,2020.In these cases,362 cases were finally included in our study,which clinically confirmed as cerebral infarction.First,the infarction ranges of these cases were manually labeled by two experienced radiologists.Then,these cases were randomly allocated to the training set(n=288),validation set(n=37),and test set(n=37),and were trained by 2D U-net model.Meanwhile,the prediction results were automatically output to the structured report of the patient.Finally,the infarction volume,diameter and CT value,which predicted by 2D U-net model,were compared with the manual labeled results,by using Bland-Altman plot.Besides,the dice similarity coefficient(DSC),volume similarity(VS),and Hausdorff distance(HD)were used to evaluate the efficacy of the model.Results:The Bland-Altman plot showed the results predicted by 2D U-net model were mostly consistent with manual labeled results,and 2.8% to 11.1% data within the 95% limits of agreement(95%LoA).Besides,the average DSC,VS,and HD in the test set were 0.66(95%CI:0.60~0.72),0.75(95%CI:0.69~0.82),and 39.69mm(95%CI:32.38~47.01),respectively.Conclusion:It is feasible to segment cerebral infarction on NCHCT by using deep learning algorithms.And the automatic structured reporting could be clinically used for the triage of patients with stroke.

关 键 词：深度学习 体层摄影术 X线计算机 脑梗死 研究报告 

分 类 号：R-056[医药卫生] R814.4