Chen Yang, Yu Hong, Fan Bin, Wang Yong, Wen Zhibo, Hou Zhihui, Yu Jihong, Wang Haiping, Tang Zhe, Li Ning, Jiang Peng, Wang Yang, Yin Weihua, Lu Bin
Department of Radiology, State Key Lab and National Center for Cardiovascular Diseases, Fuwai Hospital, National Center for Cardiovascular Diseases, Chinese Academy of Medical Sciences and Peking Union Medical College, #167 Bei-Li-Shi Street, Xi-Cheng District, Beijing, 100037, China.
Department of Radiology, Medical Imaging Center of Guizhou Province, Affiliated Hospital of Zunyi Medical University, Zunyi, Guizhou, China.
Int J Cardiovasc Imaging. 2025 May;41(5):979-989. doi: 10.1007/s10554-025-03383-0. Epub 2025 Mar 29.
To validate a fully automated, deep learning model based on coronary computed tomography angiography (CCTA) for the diagnosis of obstructive coronary artery disease (CAD) with stenosis ≥ 50%, which is commonly used as a clinical threshold for further testing and management. This model aims to improve diagnostic efficiency by automating the identification of significant coronary stenosis(≥ 50%).
This multicenter clinical trial included patients been undergone CCTA from October 13, 2022, to February 28, 2023. CCTA data from suspected coronary artery disease (CAD) patients were retrospectively analyzed using deep learning-based software for comprehensive assessment, including coronary segmentation, lumen, and stenosis determination with comparison to the reference standard of consensus by three experts. This study utilized a multi-stage deep learning framework for coronary artery segmentation and stenosis analysis from CCTA images, consisting of several key components, including the 3D Multi-resolution Cascade Convolutional Neural Network (CNN), 3D Cascade-Locally Optimized Network, and Stenosis Analysis Network. The clinical trial registry number was NCT06172985.
A total of 1090 patients (mean age: 59.90 ± 11.51 years, 47.3% female) were included in this multicenter study. Artificial intelligence (AI) demonstrated excellent performance at the patient level, accurately diagnosing ≥ 50% stenosis by assessing each patient's coronary artery condition. The AI system showed high values for accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). The values of the above statistics were 92.8%, 95.3%, 91.4%, 85.6%, and 97.3%, respectively. Excellent agreement was seen between expert readers and deep learning-determined maximal diameter stenosis for per-patient (kappa coefficients: 0.84, 95%CI: 0.81-0.88). Regarding diagnostic efficiency, comparing the AI with expert readers, the average reading time decreased from 5.94 min to 2.01 min (p < 0.001).
A novel AI-based assessment of CCTA can accurately and rapidly identify patients with coronary artery stenosis ≥ 50%, aiding in effective triage within the defined study population.
验证基于冠状动脉计算机断层扫描血管造影（CCTA）的全自动深度学习模型，用于诊断狭窄≥50%的阻塞性冠状动脉疾病（CAD），这一狭窄程度通常被用作进一步检查和治疗的临床阈值。该模型旨在通过自动识别显著冠状动脉狭窄（≥50%）来提高诊断效率。
这项多中心临床试验纳入了2022年10月13日至2023年2月28日期间接受CCTA检查的患者。对疑似冠状动脉疾病（CAD）患者的CCTA数据进行回顾性分析，使用基于深度学习的软件进行全面评估，包括冠状动脉分割、管腔和狭窄判定，并与三位专家的共识参考标准进行比较。本研究利用多阶段深度学习框架对CCTA图像进行冠状动脉分割和狭窄分析，该框架由几个关键组件组成，包括3D多分辨率级联卷积神经网络（CNN）、3D级联局部优化网络和狭窄分析网络。临床试验注册号为NCT06172985。
这项多中心研究共纳入1090例患者（平均年龄：59.90±11.51岁，47.3%为女性）。人工智能（AI）在患者层面表现出优异的性能，通过评估每位患者的冠状动脉状况准确诊断出≥50%的狭窄。AI系统在准确性、敏感性、特异性、阳性预测值（PPV）和阴性预测值（NPV）方面表现出较高的值。上述统计值分别为92.8%、95.3%、91.4%、85.6%和97.3%。专家读者与深度学习确定的每位患者最大直径狭窄之间存在高度一致性（kappa系数：0.84，95%CI：0.81 - 0.88）。在诊断效率方面，将AI与专家读者进行比较，平均阅读时间从5.94分钟降至2.01分钟（p < 0.001）。
一种基于AI的新型CCTA评估方法能够准确、快速地识别冠状动脉狭窄≥50%的患者，有助于在特定研究人群中进行有效的分流。
