基于CT的计算机视觉特征和影像组学特征的诺模图术前预测胃腺癌淋巴结转移

Nomogram Based on CT Computer Vision Features and Radiomics Features for Preoperative Prediction of Lymph Node Metastasis in Gastric Adenocarcinoma

目的:建立并验证一个基于计算机视觉特征(CVFs)和影像组学特征(RFs)的诺模图用于术前预测胃腺癌患者的淋巴结转移。方法:回顾性收集2013年7月至2018年8月经病理证实的171例GAC患者,并随机分为训练组和验证组。从每位患者的静脉期CT图像中提取CVFs和RFs。采用基于随机森林的Boruta方法进行关键特征的筛选。选定的关键CVFs和RFs分别通过logistic回归方法计算每个患者的CV-score和Rscore。使用单因素和多因素分析筛选胃腺癌淋巴结转移的独立预测因子。从分辨力、校准和临床实用性方面评估诺模图的性能。结果:分别筛选出6个关键RFs和11个关键CVFs,并计算出每个患者的R-score和CVscore。单因素和多因素分析显示R-score (OR:1.876;95%CI:1.066~3.586)和CV-score (OR:5.268;95%CI:2.672~12.240)是淋巴结转移的独立危险因素。在训练组和验证组中,诺模图的预测效能均优于单独的R-score (0.850对0.730,P=0.003;0.820对0.709,P=0.042)。诺模图和R-score对训练组中淋巴结转移概率的校准曲线显示了预测结果和观察结果之间具有良好一致性。决策曲线分析表明,诺模图获得了较高的临床净收益。结论:基于CV-score和R-score的诺模图可用于术前预测胃腺癌患者的淋巴结转移,可为临床术前个体化评估提供依据,并提高临床决策的信心。

Purpose:To establish and validate a nomogram based on computer vision features(CVFs)and radiomic features(RFs)for preoperative prediction of lymph node metastasis(LNM)in patients with gastric adenocarcinoma(GAC).Methods:A total of 171 patients with pathologically confirmed GAC between July 2013 and August 2018 were retrospectively enrolled and randomly divided into training and test cohort.CVFs and RFs were extracted from the venous phase CT images of each patient.The Boruta method based on random forest was used to select key features.The selected key CVFs and RFs were used to calculate the CV-score and R-score of each patient by logistic regression.Univariate and multivariate analysis were used to identify independent predictors of LNM in GAC patients.The nomogram was assessed in the aspects of discrimination,calibration and clinical usefulness.Results:Six key RFs and 11l key CVFs were selected,and R-score and CV-score were calculated for each patient.Univariate and multivariate analysis revealed that R-score(OR:1.876;95%CI:1.066-3.586)and CV-score(OR:5.268;95%CI:2.672-12.240)were independent risk factors for LNM.In the training and test cohort,the predictive efficacy of the nomogram was superior to the R-score(0.850 vs 0.730,P=0.003;0.820 vs 0.709,P=0.042,respectively).The calibration curves of the nomogram and R-score for the LNM probability in the training group showed a good consistency between the predicted results and the observed results.The decision curve analysis showed that the nomogram obtained a higher net clinical benefit.Conclusions:The nomogram integrated CV-score with R-score can be used to preoperatively predict for LNM in GAC patients,and can be used as a basis for individualized evaluation before operation.