CT重建算法对基于深度学习非门控冠脉钙化积分模型评价心血管风险分层的影响  

Influence of CT reconstruction kernel on the deep learning-based non-gated coronary calcium scoring model in diagnosing cardiovascular risk classification

作　　者：孙安 朱一白 史亮 萧毅 刘士远 SUN An;ZHU Yi-bai;SHI Liang(Department of Radiology,Changzheng Hospital,Naval Medical University,Shanghai 200003,China;不详)

机构地区：[1]海军军医大学长征医院影像科,上海200003 [2]海军军医大学东方肝胆医院医疗保障中心医学工程科,上海200438 [3]数坤(北京)网络科技有限公司,北京100102

出　　处：《放射学实践》2023年第4期419-425,共7页Radiologic Practice

基　　金：国家自然科学基金面上项目(82271994);国家自然科学基金面上项目(81871405);申康能力提升项目(SHDC22022310-B);军委面上项目(22BJZ07);国家卫生健康委放射影像数据库建设项目(YXFSC2022JJSJ010);上海长征医院青年启动基金(2022QN091)。

摘　　要：目的:以门控CT为标准,探讨基于深度学习的非门控冠状动脉钙化积分(DL-CACS)模型在不同CT重建算法下对心血管风险分类效能。方法:回顾性将在本院同时接受门控心脏CT和非门控低剂量胸部CT(LDCT)检查的549例患者纳入本研究。根据扫描方式(心电门控和非门控),将所有患者的图像资料分为A、B两组。对B组图像数据分别使用smooth、standard及sharp算法进行重建(作为B1、B2和B3组),并导入DL-CACS模型进行分析,获得CACS及心血管风险分类结果。以医师基于A组图像手工测量的CACS为标准,采用符合率、Bland-Altman法及组内相关系数(ICC)对3种CT重建算法下获得的DL-CACS进行分析。依据CACS(0、1~99、100~400和>400)将患者的心血管风险分为4个标准类别(1~4类,分别对应无、低、中和高风险),利用Kappa检验、受试者工作特征(ROC)曲线下面积(AUC)比较不同重建算法下DL-CACS与标准CACS对患者心血管风险分层的差异。结果:B1、B2和B3组的DL-CACS与A组之间的一致性均较好,其中以B1组最好[ICC=0.955(95%CI:0.947~0.962)]。B1、B2和B3组中模型所获得的心血管风险分层与A组之间的一致性均较好,Kappa值分别为0.839、0.827和0.770(P均<0.001),其中B1组评估高危患者的AUC最高(AUC=0.995,P<0.001)。Bland-Altman图(A组分别与B1、B2和B3组的CACS进行配对比较)显示,B1组与A组之间CACS平均差值为-0.173(95%CI:-1.748~1.402),B1组CACS超出95%一致性界限的患者数最少。结论:非门控DL-CACS模型在不同CT重建算法下均能准确地评估CACS及风险分层,而在LDCT时选择smooth重建算法,能最大程度地提高对冠脉钙化程度的评估准确性。Objective:The purpose of this study was to investigate the cardiovascular risk classification performance of the deep learning-based non-gated coronary artery calcium scoring(DL-CACS)model under different CT reconstruction kernels using ECG-gated CT as the standard.Methods:549 patients who received both ECG-gated CT and non-gated low-dose chest CT(LDCT)in our hospital were retrospectively collected.All subjects were divided into two groups:group A was with ECG-gated trigger scan,and group B was with non-gated.The images in group B were reconstructed using smooth,standard and sharp kernel(named as B1,B2 and B3 sub-groups),and imported into the DL-CACS model for analysis,and the results of CACS and cardiovascular risk classification were recorded.Taking the CACS of group A measured manually as the standard,DL-CACS obtained from different CT reconstructed kernels were analyzed using coincidence rate,ICC and Bland-Altman map.According to CACS(0,1~99,100~400,and>400),the patients'cardiovascular risk was divided into four stan-dard categories(1~4 grade,corresponding to none,low,moderate and high risk).Kappa test and area under receiver operating characteristic(ROC)curve(AUC)were used to compare the difference between DL-CACS and standard CACS in cardiovascular risk stratification of the patients under different reconstruction algorithms.Results:DL-CACS in groups B1,B2 and B3 were consistent with those in group A,among which group B1 was the best(ICC=0.955,95%CI:0.947~0.962).The cardiovascular risk classification obtained by the model in groups B1,B2 and B3 showed great consistency with that in group A,with Kappa values of 0.839,0.827 and 0.770(all P<0.001),among which the AUC of high-risk patients assessed in group B1 was the highest(AUC=0.995,P<0.001).Meanwhile,Bland-Altman plot(comparison of CACS in group A with those in groups B1,B2 and B3,respectively)showed that the mean difference of CACS between group B1 and group A was-0.173(95%CI:-1.748~1.402),and the number of patients with CACS exceeding the 95%congenial bou

关 键 词：冠状动脉钙化 体层摄影术 X线计算机 深度学习 心血管风险分层 

分 类 号：R814.42[医药卫生—影像医学与核医学] R543.3[医药卫生—放射医学] R541.4[医药卫生—临床医学]