能谱CT定量参数联合动脉期碘基图影像组学特征在肺鳞癌与肺腺癌鉴别诊断中的价值

Value of Quantitative Parameters of Spectral CT Combined with Arterial-phase Iodine-based Imaging Radiomics Feature the Differential Diagnosis for Lung Squamous Cell Carcinoma and Lung Adenocarcinoma

目的基于能谱电子计算机断层扫描(computed tomography,CT)定量参数联合动脉期碘基图影像组学特征构建诺模图模型,探讨能谱CT在肺鳞癌(lung squamous cell carcinoma,LUSC)与肺腺癌(lung adenocarcinoma,LUAD)鉴别诊断中的临床应用价值。方法收集术前行能谱CT增强检查、经病理活检或手术证实的肺癌患者89例,其中LUAD组38例,LUSC组51例。分析LUAD组和LUSC组患者在60keV动脉期单能量图像中病灶水浓度、碘浓度(iodine concentration,IC)、标准化碘浓度(normalizediodineconcentration,NIC)、CT值、能谱曲线斜率(K)、有效原子序数(effective-Z,Eff-Z)及标准化有效原子序数(normalized effective-Z,NEff-Z)。采用Logistic回归分析构建LUAD和LUSC基于能谱CT特征的鉴别诊断模型。应用ITKSnap软件提取动脉期碘基图影像组学特征;采用组内相关系数(intraclass correlation coemcient,ICC)、递归特征消除(recursive featureelimination,RFE)和最小绝对收缩和选择算子(least absolute shrinkage and selection operator,LASSO)对影像组学特征进行降维、筛选,采用Logistic回归构建基于影像组学特征的诊断模型,并计算该模型的影像组学评分(radiomics score,Rad-score),以多因素Logistic回归分析筛选出的能谱CT定量参数与Rad-score构建联合模型,并绘制诺模图。应用受试者工作特征(receiveroperatingcharacteristic,ROC)曲线、Delong检验、校正曲线、Hosmer-Lemeshow检验及临床决策曲线(decision curve analysis,DCA)对能谱CT模型、影像组学模型和诺莫图模型进行效能评价。结果LUAD组和LUSC组患者IC、NIC、CT值、K及NEff-Z比较差异均有统计学意义(P均<0.05)。多因素Logistic回归分析结果显示,IC、NIC及NEff-Z为独立影响因素,基于此构建能谱CT模型预测LUAD和LUSC的效能曲线下面积(areaunderthecurve,AUC)为0.768,准确率、灵敏度、特异度分别为70.73%、76.92%和67.86%;影像组学特征经降维后共筛选出的有意义特征有5个,一阶特征2个、二阶特征2个和形状特征1个,影像组学模型预测LUAD和LUSC的效能AUC为0.848,准确率、灵敏度、特异度分别为80.50%、83.33%和75.00%;诺莫图模型预测LUAD和LUSC的效能AUC为0.912,准确率、灵敏度、特异度分别为85.00%、92.31%和85.71%。经Delong检验显示,诺莫图模型AUC均明显高于影像组学模型和能谱CT模型(P均<0.05)。Hosmer-Lemeshow检验结果显示,能谱CT模型、影像组学模型及诺莫图模型的拟合度均良好(χ^(2)值分别为8.592、6.591、6.686,P值分别为0.378、0.581、0.570)。校准曲线分析显示,诺莫图模型预测LUAD和LUSC的预测概率曲线与理想曲线更接近,优于影像组学模型和能谱CT模型;DCA分析结果显示,诺莫图模型的AUC最大,均高于影像组学模型和能谱CT模型,临床净收益更高。结论基于能谱CT定量参数联合动脉期碘基图影像组学特征构建的诺模图模型在LUSC与LUAD鉴别诊断中具有潜在应用价值。

Objective To construct nomogram model based on the quantitative parameters of spectral computed tomography(CT)combined with the radiomics features of arterial-phase iodine-based imaging,and investigate the clinical value of spectral CT in the differential diagnosis of lung squamous cell carcinoma(LUSC)and lung adenocarcinoma(LU-AD).Methods A total of 89 patients with lung cancer underwent enhanced spectral CT examination,confirmed by pathological biopsy or surgery before surgery were enrolled in this study,including 38 cases in LUAD group and 51 cases in LUSC group.The water concentration of lesions,iodine concentration(IC),normalized iodine con centration(NIC),CT value,slope of the energy spectrum curve(K),effectiveZ(Eff-Z)and normalized effective-Z(NEff-Z)in 60 keV arterial phase monochromatic images of patients in LUAD group and LUSC group were analyzed.The differential diagnosis model for LUAD and LUSC based on spectral CT features were constructed by Logistic regression analysis.The arterial-phase iodine-based imaging features were extracted by ITK Snap software;intraclass correlation coemcient(ICC),recursive featureelimination(RFE)and least absolute shrinkage and selection operator(LASSO)were used to reduce the dimension and select the radiomics features,the diagnostic model based on the radiomics features were constructed by Logistic regression,and radiomics score(Rad-score)of the model was calculated,a joint model between the selected spectral CT quantitative parameters and Rad-score was established by multivariate Logistic regression analysis and the nomogram was drawn.The efficacy of spectral CT model,radiomics model and nomogram model were evaluated by receiver operating characteristic(ROC)curve,Delong test,calibration curve,Hosmer-Lemeshow test and decision curve analysis(DCA).Results There were significant differences in IC,NIC,CT value,K and NEff-Z between LUAD group and LUSC group(all P<0.05).Multivariate Logistic regression analysis showed that IC,NIC and NEff-Z were independent influencing factors,the area under curve(AUC)of the spectral CT model for predicting LUAD and LUSC was 0.768,the accuracy,sensitivity and specificity were 70.73%,76.92%and 67.86%respectively;a total of 5 meaningful features,2 first-order features,2 second-order features and 1 shape feature,the efficacy AUC of the imaging histology model to predict LUAD and LUSC was 0.848,with an accuracy,sensitivity and specificity of 80.50%,83.33%and 75.00%respectively;the nomogram model predicted LUAD and LUSC with an efficacy AUC of 0.912 and accuracy,sensitivity and specificity of 85.00%,92.31%and 85.71%respectively.The Delong test showed that the AUC values of the nomogram model were all significantly higher than those of the radiomics model and the spectral CT model(all P<O.05).The results of the Hosmer-Lemeshow test showed that the spectral CT model,the radiomics model and the nomogram model were all well fitted(χ^(2)were 8.592,6.591 and 6.686 respectively,P were 0.378,0.581 and 0.570 respectively).Calibration curve analysis showed that the predicting probability curves of LUAD and LUSC predicted by the nomogram model were closer to the ideal curves and better than those of the radiomics model and spectral CT model;DCA analysis showed that the AUC of the nomogram model was the largest,which were both higher than those of the radiomics model and the spectral CT model,and the net clinical benefit was higher.Conclusion Constructing a nomogram model based on quantitative parameters of spectral CT combined with the imaging radiomics features of arterial-phase iodine-based imaging has potential application in the differential diagnosis of LUSC and LUAD.