基于增强CT影像组学评分和临床分期的列线图预测胃癌脉管浸润的价值

目的 探讨基于增强CT影像组学评分（Radscore）和TNM分期的列线图预测胃癌脉管浸润（LVI）的价值。 方法 回顾性收集160例术前行上腹部CT增强检查且行术后胃癌LVI状态评估的病人,男109例,女51例,平均年龄（62.23±10.74）岁。160例病人（包括LVI阴性者92例,阳性者68例）按照7∶3比例随机分为训练集（112例）和测试集（48例）;其中,训练集中LVI阴性者60例、阳性者52例,测试集中LVI阴性者32例、阳性者16例。基于增强CT影像提取并筛选影像组学特征,建立影像组学标签并计算Radscore。采用t检验、Mann-Whitney U检验、卡方检验或Kruskal-Wallis H 检验比较LVI阳性组和阴性组间临床病理特征病人性别、年龄、肿瘤直径、TNM分期、AJCC分期、肿瘤分化程度及癌胚抗原（CEA）、糖类抗原199（CA199）]的差异,将差异有统计学意义的特征和影像组学标签纳入多因素logistic回归,建立临床影像联合模型和列线图。采用受试者操作特征（ROC）曲线评估影像组学模型和列线图的预测效能并计算相应的曲线下面积（AUC）。采用决策曲线评价影像组学模型和列线图的临床净获益。分别基于训练集及测试集中的数据绘制校准曲线对列线图进行验证。 结果 LVI阳性组和阴性组间肿瘤T分期、N分期、AJCC分期的差异均有统计学意义（均P<0.05）,且LVI阳性组的Radscore高于阴性组（P<0.05）。在测试集中,基于T分期、N分期、AJCC分期和Radscore的临床影像联合模型预测LVI的AUC值、准确度和特异度较影像组学模型分别提高了8.2%、18.2%和21.9%。决策曲线分析显示应用联合模型的临床净获益优于影像组学模型。联合模型的列线图显示Radscore得分最高,其次是AJCC分期,最后是N分期和T分期。训练集和测试集中的校准曲线显示列线图的预测结果与真实结果具有较好的一致性。 结论 联合T分期、N分期、AJCC分期和增强CT的Radscore建立的列线图能够成功预测胃癌LVI。

A nomogram based on contrast-enhanced CT radscore and TNM stage for the prediction of lymphovascular invasion in gastric cancer

Objective To investigate the value of nomogram based on contrast-enhanced CT radscore(Radscore) and TNM stage in predicting lymphovascular invasion (LVI) in gastric cancer. Methods The data of 160 patients with gastric cancer who had undergone contrast-enhanced CT of the upper abdomen and postoperative LVI assessment were retrospectively collected, including 109 men and 51 women, with a mean age of 62.23±10.74 years. A total of 160 patients, including 92 patients without LVI and 68 patients with LVI. All patients were randomly divided into a training set (consisting of 112 patients) and a testing set (consisting of 48 patients) at a rate of 7∶3. There were 60 patients without LVI and 52 patients with LVI in the training set, and 32 patients without LVI and 16 patients with LVI in the testing set, respectively. A list of radiomics features were extracted from the contrast-enhanced CT images and were selected to develop the radiomics signatures. A Radscore was then calculated for each patient. The t-test, Mann-Whitney U test, chi-squared test, and Kruskal-Wallis H test were used to compare the clinicopathologic features (including age, gender, tumor diameter, tumor node metastasis stage, AJCC stage, tumor differentiation, CEA, and CA 199) between the LVI positive group and LVI negative group. Those features with statistical significance and radiomics signatures were then fed into a logistic regression to develop a combined model and nomogram. Receiver operating characteristic (ROC) curves were used to determine the performance of the radiomics model and nomogram, and the area under the curves (AUCs) were calculated. The decision curve was conducted to determine the clinical usefulness of the radiomics model and nomogram. Calibration curves were plotted to assess the calibration of the nomogram on the training and testing data sets. Results T stage, N stage, AJCC stage, and Radscore were significantly different between the LVI positive group and LVI negative group (all P<0.05). The Radscore was significantly higher in the LVI positive group than in the LVI negative group (P<0.05). Compared to the radiomics model, the AUCs, accuracy, and specificity of the combined model based on the combination of T stage, N stage, AJCC stage, and Radscore were increased 8.2%,18.2%, and 21.9%, respectively, in the testing set. The decision curve showed the combined model performed better than the radiomics model. The nomogram showed the Radscore achieved the highest score, followed by the AJCC stage, N stage and T stage. The calibration curve of the nomogram for the probability of LVI demonstrated relatively good agreement between prediction and observation in the training and testing data sets. Conclusion The nomogram based on the combination of T stage, N stage, AJCC stage, and contrast-enhanced CT Radscore could predict LVI successfully.