基于冠状动脉钙化斑块伪影特征识别的非门控CT钙化积分定量研究

Approximately 50% of coronary artery disease (CAD) patients are asymptomatic. Coronary artery calcium scoring is a reliable method to diagnose and stratify the cardiovascular risk in asymptomatic CAD patients. In the tens of millions of subjects who were examined by thoracic CT in China annually, a large proportion of them are potential population for asymptomatic CAD. If thoracic CT can evaluate calcium score, the high cardiovascular risk patients would be screened out effectively. However, due to coronary artery motion artifacts, the calcium scoring in nontriggered thoracic CT is not accurate enough. The preliminary study simulated two-dimensional(2D) linear motion of calcified plaques using a robotic arm, and showed the pattern of motion artifacts, which associated with plaque density, size and velocity. We hypothesized that accurate calcium scoring can be reached by regression based on characteristics of motion artifacts in nontriggered thoracic CT. This study aims to simulate three-dimensional(3D) motion of calcified plaques using a robotic arm to obtain systematic and modeling data on artifacts, and prospectively include 300 human subjects to acquire practical data. After quantification of multiple artifact characteristics on these data, a new regression algorithm to calculate accurate score will be established by artificial neural network method. Thus, accurate scoring using nontriggered thoracic CT will be achieved. Finally, 200 subjects will be prospectively included to validate the new algorithm. This study will provide theory and examination technique to easily and reliably screen out high cardiovascular risk patients from thoracic CT subjects.

约50%的冠心病为无症状型，诊断和心血管危险度分层的一种可靠方法是CT冠脉钙化积分。我国每年进行数千万例胸部CT检查，被检人群中相当数量是潜在无症状冠心病患者。如果胸部CT能测量钙化积分将能有效检出心血管高危患者，但非心电门控胸部CT测量的积分受冠脉运动伪影影响不够准确。前期实验通过用机械臂模拟钙化斑块二维线性运动，发现伪影特征同斑块大小、密度和速度等有相关性。我们设想可以从钙化斑块运动伪影特征回归分析得出准确的积分。本课题拟使用机械臂模拟冠脉三维运动获得系统化、理论模型化的运动伪影数据，前瞻性纳入300例患者获得实际条件下的伪影数据。然后量化伪影的多种特征，用人工神经网络方法根据伪影特征回归分析得到准确的钙化积分，解决胸部CT测量冠脉钙化积分的准确性问题。最后纳入200例患者评估所开发新算法和技术的可靠性。本研究将为在临床应用中简捷、有效地检出心血管高危患者提供理论基础和技术保证。

约50%的冠心病为无症状型，诊断和心血管危险度分层的一种可靠方法是CT冠脉钙化积分。我国每年进行数千万例胸部CT 检查，被检人群中相当数量是潜在无症状冠心病患者。如果胸部CT 能测量钙化积分将能有效检出心血管高危患者，但非心电门控胸部CT测量的积分受冠脉运动伪影影响不够准确。前期实验通过用机械臂模拟钙化斑块运动，发现伪影特征同斑块大小、密度和速度等有相关性。我们设想可以从钙化斑块运动伪影特征回归分析得出准确的积分。本课题使用机械臂模拟冠脉运动获得系统化、理论模型化的运动伪影CT图像数据。然后用深度卷积神经网络(CNN)处理训练组图像，使用迁移学习方法，基于大数据为基础的模型进行微调，训练模型用于根据伪影特征对CT图像进行分类。如果CNN能根据伪影识别正确的钙化斑块，则使用该斑块静止状态下的钙化积分为校正积分。使用8重折叠交叉验证方法，发现对所有图像分类的总准确率为78.2±6.1%，对高、中和低密度斑块分类的准确率分别为87.9±4.9%, 74.1±6.4% 和72.7±5.0%。与静止状态下的Agatston积分相比较，运动状态下的积分变异性为37.8% (第1、3四分位: 10.5%, 68.8%)。使用CNN校正后，变异性大幅缩小至3.7% (1.9%, 9.1%) (p<0.001)。与传统上使用CT值>130HU定义钙化斑块比较，CNN将运动状态下钙化斑块的检出率从65%提高到85%。总之，在这个基础性研究中，CNN显示了识别非门控CT钙化斑块伪影，并进行分类的潜力。CNN校正能大幅度降低钙化积分测量变异性，提高对钙化斑块检出的敏感性。本研究为在临床应用中使用非门控CT评估钙化积分，进而有效地检出心血管高危患者提供理论基础和技术保证。