目的 观察冠状动脉CT血管成像（CCTA）影像组学机器学习（ML）模型联合冠状动脉周围脂肪衰减指数（FAI）预测冠状动脉斑块进展的价值。方法 回顾性收集194例CCTA显示冠状动脉斑块并接受至少1次CCTA复查患者,基于首次、末次CCTA计算斑块负荷（TPB）年变化值（△TPB/y）分为进展组（△TPB/y≥中位△TPB/y）与无进展组（△TPB/y＜中位△TPB/y）,并按8∶2比例划分训练集（n=155）与验证集（n=39）。以单因素及多因素logistic回归分析筛选斑块进展的临床及首次CCTA相关影响因素,以之构建CCTA模型。基于训练集首次CCTA提取及筛选斑块最佳影像组学特征,分别采用随机森林（RF）、高斯过程（GP）、偏最小二乘判别分析（PLS-DA）、二次判别分析（QDA）及支持向量机（SVM）算法构建ML模型,于验证集验证模型效能并选择最优ML模型,以之联合CCTA模型构建联合模型;评估各模型预测冠状动脉斑块进展效能。结果 194例中,进展组97例、无进展组97例。训练集155例中,77例斑块进展、78例斑块无进展;验证集39例中,20例斑块进展、19例斑块无进展。FAI是冠状动脉斑块进展的独立预测因素（OR=1.08,P＜0.001）,以之构建CCTA模型。基于训练集数据筛选的10个最佳影像组学特征分别构建RF、GP、PLS-DA、QDA、SVM模型,RF模型在训练集和验证集的曲线下面积（AUC）均较高,为最优ML模型。CCTA模型、RF模型及联合模型在训练集的AUC分别为0.684、0.847及0.861,在验证集的AUC分别为0.629、0.768及0.821。结论 CCTA影像组学ML模型联合FAI能有效预测冠状动脉斑块进展。
Objective To evaluate the value of coronary CT angiography （CCTA） radiomics machine learning （ML） model combined with pericoronary fat attenuation index （FAI） for predicting coronary plaques progression. Methods Totally 194 patients with CCTA showing coronary plaques and received at least one CCTA review afterwards were retrospectively collected. The annual change value of total plaque burden （△TPB/y） was calculated based on the first and last CCTA to assess plaque progression. All patients were categorized into non-progressive （△TPB/y＜median △TPB/y） and progressive （△TPB/y≥median △TPB/y） groups. The patients were divided into training set （n=155） and validation set （n=39） at the ratio of 8∶2.Univariate and multivariate logistic regression analyses were used to screen clinical and primary CCTA related factors for plaque progression, and CCTA model was constructed. Radiomics features were extracted and screened based on primary CCTA to build ML models using random forest （RF）, Gaussian process （GP）, partial least squares discriminant analysis （PLS-DA）, quadratic discriminant analysis （QDA） and support vector machine （SVM） algorithms. The effectiveness of all models was verified in validation set and the optimal ML model was selected. And its combination with CCTA model constructed combined model. The efficacy of each model for predicting coronary plaques progression was evaluated. Results Of 194 cases, 97 were in progressive group and 97 were in non-progressive group. The training set included 77 cases of plaques progression and 78 of plaques non-progression, and the validation set included 20 of plaques progression and 19 of plaques non-progression. FAI was the independent predictor of plaque progression （OR=1.08, P＜0.001） and CCTA model was constructed. Ten optimal radiomics features based on training set were selected to build RF, GP, PLS-DA, QDA and SVM models. The area under the curve （AUC） of RF model in training set and validation set were both high, was considered as the optimal ML model. The AUC of CCTA, RF and combined models in training set was 0.684, 0.847 and 0.861, respectively, while was 0.629, 0.768 and 0.821 in validation set, respectively. Conclusion CCTA radiomics ML model combined with FAI could effectively predict coronary plaques progression.
知识与财富的链接
