基于冠状动脉周围脂肪组织影像组学特征及神经网络模型预测冠状动脉狭窄血流动力学严重程度

目的 评估基于神经网络方法构建的预测模型能否精准评估冠状动脉狭窄的血流动力学严重程度（缺血或不缺血）。 方法 回顾性收集行冠状动脉CT血管成像（CCTA）及有创冠状动脉造影的血流储备分数（FFR）测量的92例冠状动脉疾病病人的临床及影像资料,其中男66例,女26例;平均年龄（58.3±10.3）岁。共纳入122支冠状动脉血管。依据FFR值将122支冠状动脉血管分为2组,即狭窄组（FFR≤0.8,68支）和非狭窄组（FFR>0.8,54支）。基于CCTA影像数据,选取冠状动脉周围脂肪组织（PCAT）区域的468个影像组学特征进行分析。构建3种冠状动脉狭窄预测模型,包括神经网络模型、传统统计学模型和最小绝对值收敛与选择算子模型。采用受试者操作特征曲线下面积（AUC）评估3种模型预测冠状动脉狭窄的性能。采用Pearson相关分析神经网络特征、原始影像组学特征与真实标签的相关性。采用独立样本t检验比较2组的影像组学特征。 结果 3种预测模型中,神经网络模型的预测效能最高,其准确度、敏感度、特异度和AUC分别为81.19%、81.23%、81.16%和0.781 3（0.773 8~0.788 8）。神经网络特征与真实冠状动脉狭窄标签的相关性最大绝对相关系数（r_最大）=0.683 8,P<0.001,平均绝对相关系数（r_平均）=0.261 1]高于原始影像组学特征与真实标签的相关性（r_最大=0.238 9,P=0.008和r_平均=0.090 5）。狭窄组的W6_surface_area高于非狭窄组,而W6_Auto Correlation_mean低于非狭窄组（均P<0.05）,其余特征差异均无统计学意义（均P>0.05）。 结论 以影像组学特征为输入的神经网络模型可以很好地预测冠状动脉狭窄,其中10个PCAT区域影像组学特征或许在预测冠状动脉狭窄的血流动力学方面具有重要意义。

Predict hemodynamic severity of coronary artery stenosis with peri-coronary adipose tissue based radiomics and neural network model

Objective To evaluate whether a neural network based prediction model can assess the hemodynamic severity of coronary artery stenosis (ischemia or non- ischemia) precisely. Method Ninety-two patients with coronary artery disease, who underwent coronary computed tomography angiography (CCTA) examination, invasive coronary angiography examination, and fractional flow reserve examination, were recruited, including 66 males and 26 females, average age 58.3±10.3 years. Totally, 122 coronary arteries were evaluated in this study and were split into stenosis group (FFR≤0.8, 68 coronary arteries) and non-stenosis group (FFR>0.8, 54 coronary arteries). Based on CCTA imaging data, 468 radiomics features from peri-coronary adipose tissue (PCAT) area were extracted. Three coronary artery stenosis predicting models were constructed, including the neural network model, traditional statistical model, and least absolute shrinkage and selection operator model. The area under the receiver operating characteristic curve (AUC) was used to evaluate the predicting performance of the models. Pearson correlation was used to analyze the relationship between neural network or radiomics features and real stenosis label. Independent sample t test was used for the comparison between the two groups. Results In these three models, the neural network model had the highest predicting performance, with the accuracy, sensitivity, specificity, and AUC of 81.19%, 81.23%, 81.16%, and 0.781 3 (0.773 8-0.788 8), respectively. The correlation between neural network features and real label of coronary artery stenosis maximum absolute correlation coefficient (r_max)=0.6838, P<0.001, and average absolute correlation coefficient (r_mean)=0.261 1] was higher than the one between radiomics feature and label (r_max=0.238 9, P=0.008, and r_mean=0.090 5). The W6_surface_area in the stenosis group was higher than in the non-stenosis group, while W6_Auto Correlation_mean was lower than in the non-stenosis group (all P<0.05). Other features had no statistically significant between two groups(all P>0.05). Conclusion With radiomics features as input, neural network has a good performance on coronary artery stenosis prediction. Among them, 10 PCAT area radiomics features may be very important for the prediction of the hemodynamic severity of coronary artery stenosis.