摘要
目的探讨基于冠状动脉CT血管成像的深度学习(DL)模型评估冠心病管腔狭窄的诊断性能。资料与方法回顾性分析2014年7月—2020年7月北京大学首钢医院89例疑似冠心病患者的冠状动脉CT血管成像资料,采用侵入性冠状动脉造影作为参考标准,评价DL模型对冠心病管腔狭窄的诊断性能。管腔直径狭窄≥50%认为是梗阻性冠状动脉狭窄,并在斑块类型、斑块长度、斑块累及血管水平,应用受试者工作特征曲线下面积(AUC)比较DL模型和医师的诊断效能,同时计算敏感度、特异度、阳性预测值、阴性预测值和准确度。结果DL模型诊断阻塞性冠状动脉狭窄的AUC为0.92,敏感度为86.2%,特异度为87.6%,阳性预测值为66.37%,阴性预测值为95.71%,准确度为87.28%。在斑块类型水平,DL模型对混合型斑块所致管腔狭窄的敏感度最高(100%),对非钙化斑块所致管腔狭窄的特异度最高(88.9%),对非钙化斑块所致管腔狭窄的整体诊断效能最优,准确度为89.43%,AUC为0.94。在斑块长度水平,DL模型对局限性斑块与节段性斑块所致管腔狭窄的诊断效能接近,AUC分别为0.91和0.95。在斑块累及血管水平,DL模型对左主干、右冠状动脉、前降支、回旋支及分支血管的管腔狭窄诊断AUC分别为1.00、0.96、0.90、0.90和0.92。DL模型对不同长度斑块所致管腔狭窄的诊断性能差异有统计学意义(χ^(2)=8.43,P=0.01),对不同类型斑块所致管腔狭窄的诊断性能差异无统计学意义(χ^(2)=0.77,P=0.68),对不同累及血管管腔狭窄的诊断性能差异无统计学意义(χ^(2)=9.43,P=0.05)。对于非钙化斑块、混合斑块及节段性斑块所致管腔狭窄,DL模型的诊断性能高于医师,差异有统计学意义(Z=2.53、2.52、2.49,P=0.01)。结论基于冠状动脉CT血管成像的DL模型诊断冠心病具有较高的准确性,是诊断冠心病的可靠辅助工具。
Purpose To investigate the diagnostic performance of deep learning(DL)model based on coronary artery CT angiography in evaluating coronary artery stenosis.Materials and Methods The coronary artery CT angiography data of 89 patients with suspected coronary heart disease were retrospectively analyzed from July 2014 to July 2020 in Peking University Shougang Hospital.The diagnostic performance of DL model for coronary artery stenosis was evaluated by using invasive coronary angiography as reference standard.Lumen diameter stenosis≥50%was considered obstructive.The diagnostic performances between DL model and doctors were compared by the areas under the receiver operating characteristic curve(AUC)of plaque type,plaque length and vessels involved levels,the sensitivity,specificity,positive predictive value,negative predictive value and accuracy were further calculated.Results AUC of 0.92 was obtained by DL model to diagnosis obstructive coronary artery stenosis,and the sensitivity,specificity,positive predictive value,negative predictive value and accuracy was 86.2%,87.6%,66.37%,95.71%and 87.28%,respectively.At the level of plaque type,DL model had the highest sensitivity(100%)for coronary stenosis caused by mixed plaques and the highest specificity(88.9%)for coronary stenosis caused by non-calcified plaques,and had the best overall diagnostic performance for coronary stenosis caused by non-calcified plaques with accuracy of 89.43%and AUC of 0.94.At the level of plaque length,DL model had similar diagnostic performance for coronary stenosis caused by localized plaques and segmental plaques,with AUC of 0.91 and 0.95,respectively.At the level of vessels involved,the AUC of DL model to diagnosis coronary artery stenosis of left main artery,right coronary artery,left anterior descending artery,left circumflex artery and branch vessels were 1.00,0.96,0.90,0.90 and 0.92,respectively.There was statistically significant difference in the diagnostic performance of DL model for coronary stenosis caused by plaques of different lengths(χ^(2)=8.43,P=0.01),but there was no statistically significant difference in the diagnostic performance of DL model for coronary stenosis caused by plaques of different types(χ^(2)=0.77,P=0.68),and there was no statistically significant difference in the diagnostic performance of DL model for different vessels involved stenosis(χ^(2)=9.43,P=0.05).For coronary stenosis caused by non-calcified plaques,mixed plaques and segmental plaques,the diagnostic performance of DL model was significantly higher than that of doctors,and the differences were statistically significant(Z=2.53,2.52,2.49,P=0.01).Conclusion Deep learning model based on coronary artery CT angiography has high accuracy in the diagnosis of coronary heart disease and is a reliable auxiliary tool to diagnosis coronary heart disease.
作者
耿冀
常玉莲
张滨
王思雯
张番栋
GENG Ji;CHANG Yulian;ZHANG Bin;WANG Siwen;ZHANG Fandong(Department of Radiology,Peking University Shougang Hospital,Beijing 100144,China)
出处
《中国医学影像学杂志》
CSCD
北大核心
2023年第7期706-712,共7页
Chinese Journal of Medical Imaging
作者简介
通信作者:张滨.zhangbin-m@126.com。
