基于CT 影像组学特征的机器学习模型预测早期急性胰腺炎进展的价值

Predicting progression of early acute pancreatitis by machine learning model based on CT radiomics

目的:探讨基于平扫及增强CT影像组学特征的机器学习模型预测早期急性胰腺炎(AP)进展的价值。方法:回顾性分析2013年11月~2021年1月弋矶山医院104例AP患者资料,其中进展期40例,非进展期64例。随机将患者按7∶3的比例分为训练组和验证组。对胰腺实质全部层面进行手动勾画ROI并进行三维融合,采用AK软件提取纹理特征,使用最小冗余最大相关(mRMR)和最小绝对值收敛和选择算子(LASSO)回归分析筛选出最佳纹理特征。采用多因素Logistic回归(LR)、随机森林(RF)、支持向量机(SVM)三种方法建立影像组学预测模型。利用受试者工作特征(ROC)曲线评价不同模型的预测效能。结果:基于平扫、动脉、静脉、延迟及四期联合序列构建的LR模型在训练组中的曲线下面积(AUC)分别为0.83、0.84、0.78、0.86、0.85,在验证组中分别为0.76、0.79、0.79、0.84、0.78;基于延迟期的12个最佳特征构建的RF和SVM模型在训练组中的AUC分别为0.78、0.84,在验证组中的AUC分别为0.72、0.77。结论:基于延迟期CT图像的影像组学特征构建的LR模型预测早期AP患者病情进展具有较高的价值。

Objective:To assess the value of machine learning model based on plain and enhanced CT image radiomic for predicting progression of early acute pancreatitis(AP).Methods:The image data were retrospectively analyzed in 104 cases confirmed as AP in our hospital between November 2013 and January 2021.Of the 104 cases,40 were in progression,and 64 in non-progression.All cases were randomized to training group and verification group by ration of 7:3.Manual ROI mapping was performed on all levels of pancreas on CT images,and constructed via 3D fusion.AK software was used to extract texture features.Minimum redundancy maximum correlation(mRMR),minimum absolute value convergence and selection operator(LASSO)regression analysis were used to determine the optimal texture feature.Multiple factor logistic regression(LR),random forest(RF)and support vector machine(SVM)were used to establish the prediction model of radiomics.Receiver operating characteristic(ROC)curve was developed to evaluate the prediction efficiency by diverse model.Results:AUC of LR model based on plain,arterial,venous,delayed and combined phase in the training group was 0.83,0.84,0.78,0.86 and 0.85;and in the verification group was 0.76,0.79,0.79,0.84 and 0.78,respectively.AUC for RF and SVM model based on the 12 optimal features in delayed phase was 0.78 and 0.84 in the training group,and 0.72 and 0.77 in the verification group,respectively.Conclusion:LR model of CT radiomics features generated at delayed phase has higher value for predicting progression of early acute pancreatitis.