基于深度学习的冠状动脉CTA自动后处理技术诊断冠心病的价值研究

Diagnostic value of coronary CT angiography automatic post-processing technique based on deep learning for coronary atherosclerotic heart disease

关键词:  人工智能  深度学习  冠心病  冠状动脉CT血管成像

Key words:  artificial intelligence  deep learning  coronary artery disease  coronary computed tomography angiography

摘要:

目的 探讨一种基于深度学习（deep learning, DL）的冠状动脉CT血管成像（coronary computed tomography angiography, CCTA）人工智能（artificial intelligence, AI）辅助诊断系统诊断冠心病（coronary artery disease， CAD）的临床价值。 方法 回顾性分析4周内先后完成CCTA和侵入性冠状动脉血管造影（coronary angiography， CAG）检查的69例患者，共462个冠状动脉近端血管节段。比较AI辅助诊断技术和CCTA人工分析影像数据所用时间。分别在患者水平、血管节段水平评估AI和CCTA对CAD的诊断效能；在各血管节段水平对AI、CCTA诊断结果进行误差分析。 结果 AI耗时明显低于CCTA[（5.5±0.9） min vs (14.2±1.8） min]。在患者水平和总血管节段水平，AI识别显著狭窄（狭窄程度≥50%）的灵敏度分别为97.9%、79.6%，阴性预测值分别为90.9%、93.2%，准确性分别为82.6%、80.5%；CCTA识别显著狭窄的灵敏度分别为96.6%、85.9%，阴性预测值分别为81.8%、92.9%，准确性分别为94.2%、91.6%；CCTA识别显著狭窄的准确性较高，AUC分别为0.984（95%CI： 0.9591.000）、0.960（95%CI： 0.9420.977），与CAG差异无统计学意义（均P＞0.05）；AI识别显著狭窄有一定准确性，AUC分别为0.864（95%CI： 0.7480.979）、0.823（95%CI： 0.7800.865），与CAG差异有统计学意义（均P＜0.05）。在各血管节段水平，AI检测显著狭窄的准确性分别为89.9%、78.3%、92.8%、69.6%、60.9%、88.4%、85.4%。对于钙化斑块、混合斑块、非钙化斑块、心肌桥及血管解剖异常，AI分别漏诊11、20、29、7、2段，误诊7、3、5、1、5段；CCTA分别漏诊2、5、6、2、2段，误诊9、3、4、4、2段。 结论 基于DL的CCTA自动后处理技术能高效完成CCTA数据分析，有较高的CAD诊断效能，具有成为临床诊断CAD辅助工具的潜力。而人机结合可能有望进一步提高CAD的诊断效能。

Abstract:

Objective To explore the clinical value of a novel deep learning-based coronary CT angiography(CCTA) artificial intelligence(AI)-assisted system for the diagnosis of coronary artery disease(CAD). Methods Sixty-nine patients(462 proximal coronary segments) underwent CCTA and invasive coronary angiography(CAG) examinations within 4-week interval in Tongji Hospital from February 2021 to June 2022. The imaging results were analyzed by artificial intelligence-assisted system(AI group) and manually(CCTA group), respectively. The time required for image analysis of two groups was compared. Using CAG results as gold standard, the diagnostic efficacy of AI and CCTA for vascular stenosis at patients and vascular segment levels were evaluated. Error analysis was performed on the diagnosis results of AI and CCTA at each vascular segment level. Results The time required for the AI analysis was significantly shorter than that of CCTA [(5.5±0.9) min vs (14.2±1.8) min, P<0.05]. At the patient level and vascular segment level, the sensitivity of AI to identify significant stenosis(≥50%) was 97.9% and 79.6%, the negative predictive value was 90.9% and 93.2%, and the accuracy was 82.6% and 80.5%, respectively. The sensitivity of CCTA to identify significant stenosis was 96.6% and 85.9%, the negative predictive value was 81.8% and 92.9%, and the accuracy was 94.2% and 91.6%, respectively. The area under ROC curve(AUC) of CCTA in identifying significant stenosis at patients and vascular segment levels was 0.984(95%CI： 0.959-1.000) and 0.960(95%CI： 0.942-0.977), respectively, and there were no significant differences between CCTA and CAG(P>0.05); while the AUC of AI was 0.864(95%CI： 0.748-0.979) and 0.823(95%CI： 0.780-0.865), respectively, which were significantly different from CAG(P<0.05). The accuracy values of AI detection for significant stenosis at the level of each proximal coronary vessel segment were 89.9%, 78.3%, 92.8%, 69.6%, 60.9%, 88.4%, and 85.4%. For calcified plaque, mixed plaque, non-calcified plaque, myocardial bridge and vascular anatomical abnormalities, AI missed 11, 20, 29, 7, 2 segments, and misdiagnosed 7, 3, 5, 1, 5 segments, respectively; while CCTA missed 2, 5, 6, 2, 2 segments, and misdiagnosed 9, 3, 4, 4, 2 segments, respectively. Conclusion The deep learning-based CCTA automatic post-processing technique shows high efficiency in CCTA data analysis with certain diagnostic accuracy for CAD. It may be a assistant tool for the diagnosis of CAD clinically.