基于纹理分析的肺结节良恶性鉴别诊断模型的可视化研究

The Development of Nomogram Predicting the Probability of Malignancy or Benign Nature of Pulmonary NodulesBased on Computed Tomographic Texture Analysis

目的通过CT纹理分析(CTTA)技术分析肺结节的良恶性特征,基于CTTA参数构建预测肺结节良恶性的鉴别诊断模型。方法将经病理证实的153例肺结节患者(2015年12月至2016年10月95例作为训练集,2016年10月至2017年4月58例作为测试集)纳入CTTA研究,其中结节为良性的患者69例,结节为恶性的患者84例。所有患者均在术前行CT增强扫描,对肺结节动、静脉期的图像采用FireVoxel软件进行纹理特征分析[熵(entropy)、不均匀性(inhomogenity)、峰度(kurtosis)、均值(mean)和偏度(skewness)]。根据Logistic回归分析结果构建列线图模型,并用测试集验证模型的预测价值。采用受试者工作特征(ROC)曲线分析各CTTA参数的诊断效能以及比较预测模型在训练集和测试集的差异性;最后采用决策曲线分析(DCA)法综合评价模型临床实用价值。结果动脉期entropy、inhomogenity、kurtosis、mean和skewness和静脉期entropy、mean和skewness在良恶性结节中差异有统计学意义(P值均<0.05);而静脉期inhomogenity和kurtosis两组间差异无统计学意义(P值均>0.05)。ROC曲线分析发现动脉期entropy、inhomogenity、kurtosis、mean、skewness以及静脉期entropy、mean和skewness具有一定的鉴别诊断价值(P值均<0.05)。多因素Logistic回归分析显示动脉期entropy和静脉期mean是危险因素,而动脉期skewness是保护性因素;基于多因素Logistic回归分析结果构建的列线图模型在训练集中其鉴别能力稍高于测试集[训练集曲线下面积(AUC)=0.981,95%CI:0.960~1.000,测试集AUC=0.936,95%CI:0.870~1.000],但差异无统计学意义(P=0.1919)。DCA表明,概率阈值从10%起采用CT纹理特征参数构建的肺结节良恶性的鉴别诊断模型可以使得患者净获益。结论CTTA的动脉期entropy、静脉期mean和skewness对肺结节具有较大的鉴别诊断价值,构建的预测模型具有良好的鉴别诊断能力,利用列线图模型的可视化特点,可对肺结节进行准确分类,有助于临床影像科医师对结节的良恶性加以判断。

Objective To analyze the benign and malignant characteristics of lung nodules by Computed Tomographic Texture Analysis(CTTA)technology,and construct a differential diagnosis model for predicting benign and malignant lung nodules based on CTTA parameters.Methods 153 patients with pulmonary nodules confirmed by pathology(95 cases from December 2015 to October 2016 as the training set and 58 cases from October 2016 to April 2017 as the test set)were included in the CTTA study.There were 69 patients with benign nodules and 84 patients with malignant nodules.All patients underwent enhanced CT examinations before the operation,and the images of the arterial and venous phases of the lung nodules were analyzed for texture features(entropy,inhomogenity,kurtosis,mean and skewness)using FireV oxel software.We constructed a nomogram model based on the logistic regression analysis results,and used the test set to verify the predictive value of the model.The ROC curve was used to analyze the diagnostic efficiency of each CTTA parameter and the difference between the training set and the test set of the prediction model was compared.Finally,the decision curve analysis(DCA)method was used to comprehensively evaluate the clinical practical value of the model.Results 1.Arterial phase entropy,inhomogenity,kurtosis,mean and skewness and venous stage entropy,mean and skewness were statistically different in benign and malignant nodules(P<0.05);while venous phase inhomogenity and kurtosis were different between the two groups No statistical significance(all P>0.05).2.ROC curve analysis found that the arterial phase entropy,inhomogenity,kurtosis,mean,skewness and venous phase entropy,mean and skewness have certain differential diagnostic value(P<0.05).3.Multivariate logistic regression analysis shows that the arterial phase entropy and venous phase mean are risk factors,while the arterial phase skewness is a protective factor.Thenomogram modelconstructed based on the results of multivariate logistic regression analysis had a slightly higher discriminating ability in the training set.Test set(training set AUC=0.981,95%CI:0.960~1.000),test set AUC=0.936,95%CI:0.870~1.000),but the difference was not statistically significant(P=0.1919>0.05).3.The analysis of the decision curve shows that the differential diagnostic model of benign and malignant pulmonary nodules constructed by using CT texture feature parameters with a probability threshold of 10%can bring net benefits to patients.Conclusion The arterial phase entropy,venous phase mean and skewness of CTTA have great differential diagnostic value for pulmonary nodules.The constructed predictive model has good differential diagnostic ability.Using the visualization characteristics of the nomogram model,lung nodules can be diagnosed.Accurate classification helps radiologist todistinguish between the benign and malignant nodules.