基于增强CT影像组学模型术前预测胰腺神经内分泌肿瘤病理分级

Preoperative prediction of pathological grade of pancreatic neuroendocrine tumors based on contrast-enhanced CT radiomics model

目的观察基于CT影像组学模型术前预测胰腺神经内分泌肿瘤(PNET)病理分级(G1和G2/3级)的价值。方法回顾性分析145例经病理证实的PNET,分为训练组91例、验证组54例,2组各自来源于同一医疗机构。基于训练组动脉期和门脉期CT图像提取PNET影像组学特征,以Pearson相关分析及ReliefF算法进行筛选;采用Logistic回归,针对差异有统计学意义的参数构建预测PNET病理分级的联合影像组学模型,绘制受试者工作特征(ROC)曲线,计算曲线下面积(AUC),以敏感度、特异度及准确率评估其诊断效能,并以验证组加以验证。结果基于训练组动脉期与门脉期CT图像构建的联合影像组学模型具有良好预测效能,AUC为0.86[95%CI(0.78,0.94)],截断值为0.63时,敏感度为78.95%,特异度为85.29%,准确率为81.32%。验证组预测PNET病理分级AUC为0.85[95%CI(0.75,0.95)],截断值为0.63时,敏感度为84.61%,特异度为75.00%,准确率为79.63%。结论基于增强CT图像构建的影像组学模型对于术前预测PNET病理分级具有一定价值。

Objective To explore the value of CT radiomics model for preoperative predicting pathological grade(G1 and G2/3)of pancreatic neuroendocrine tumors(PNET).Methods A total of 145 patients with pathologically confirmed PNET were included,including 91 in training group and 54 in validation group,in each group coming from one same hospital.Radiomics features of PNET were extracted based on arterial phase and portal venous phase CT images in training group.Pearson correlation analysis and ReliefF algorithm were used to select radiomics features,and Logistic regression was used to construct radiomics model for predicting pathological grade of PNET.Then receiver operating characteristic(ROC)curve was drawn,and the diagnostic performance of model was evaluated with area under the curve(AUC),accuracy,sensitivity and specificity,and were validated in validation group.Results The combined radiomics model based on arterial phase and portal venous phase CT image achieved good prediction performances.In training group,the AUC was 0.86(95%CI[0.78,0.94]),the intercept value was 0.63,the sensitivity,specificity and accuracy was 78.95%,85.29%and 81.32%,respectively.In the validation group,the AUC was 0.85(95%CI[0.75,0.95]),the intercept value was 0.63,the sensitivity,specificity and accuracy was 84.61%,75.00%and 79.63%,respectively.Conclusion The radiomics model based on contrast-enhanced CT images had certain value for preoperative prediction of pathological grade of PNET.