【作者】 李敏；

【导师】 王化；

【作者基本信息】 天津医科大学 ， 影像医学与核医学， 2022， 硕士

【摘要】 目的:探究双层探测器光谱CT影像组学特征对非小细胞肺癌(non-small cell lung cancer,NSCLC)PD-L1表达的预测价值。方法:前瞻性收集2019年9月至2021年3月在天津医科大学肿瘤医院术前行双层探测器光谱CT(简称IQon光谱CT)胸部增强扫描且经手术病理证实的NSCLC患者,根据纳入和排除标准共217例患者纳入研究。采用免疫组化方法检测患者手术切除标本的PD-L1表达,并用肿瘤比例评分(tumor proportion score,TPS)表示,以1%为阈值将其分为PD-L1阳性和阴性两组。将217例患者按7:3随机划分为训练集和测试集。收集患者的临床病理特征,包括性别、年龄、吸烟史、TNM分期及病理类型;评估CT影像学特征,包括肿瘤位置、大小、密度、形状、分叶征、毛刺征、胸膜牵拉征、血管集束征、空气支气管征、空洞/空泡征;测量IQon光谱CT定量参数,包括碘浓度(iodine concentration,IC)、标准化碘浓度(normalized iodine concentration,NIC)、能谱曲线斜率(slope rate of the HU curve,λHU)及有效原子序数(effective atomic number,Zeff)。分别从PACS系统及IQon光谱CT后处理工作站以DICOM格式保存120k Vp混合能量下的增强常规图像(contrast-enhanced conventional images,CCI)、碘密度图(iodine map,IM)及虚拟平扫(virtual non-contrast,VNC)图像。利用ITK-SNAP软件在CCI横断面对肿瘤进行逐层勾画,得到三维感兴趣区(region of interest,ROI),并复制到IM和VNC图像。使用Python软件从三个序列图像的ROI分别提取1334个影像组学特征,这些特征由一阶、二阶、形状及变换特征构成。特征筛选和建模分四步进行。第一步:筛选出观察者间相关系数ICC≥0.8的特征;第二步:采取Zscore标准化方法对组学特征进行标准化;第三步:先使用独立样本t检验或Mann-Whitney U检验筛选出与PD-L1表达相关的特征(P<0.05),再使用spearman或pearson相关系数评估特征的相关性,筛选出相关系数r<0.8的特征。第四步:采用LASSO-logistic回归建立预测PD-L1表达的模型,并计算影像组学评分(Radscore)。共建立四个模型,分别为基于CCI、IM、VNC图像以及三者联合的影像组学模型。对PD-L1阳性和阴性组间临床病理特征、影像学特征及光谱CT定量参数的差异行单因素分析,将P<0.05的变量纳入多因素logistic回归,筛选出与PD-L1表达相关的独立预测因子。将临床病理特征、CT影像学特征及光谱CT定量参数中与PD-L1相关的独立预测因子与影像组学联合模型的Radscore结合,建立预测PD-L1表达的临床-影像联合模型,并构建诺模图。所有筛选特征和建立模型的过程均在训练集中进行,在测试集中测试模型的预测效能。模型效能均采用ROC曲线进行评价,以Delong检验比较模型AUC的差异。结果:训练集共152例患者,其中PD-L1阳性67例,阴性85例。测试集共65例患者,其中PD-L1阳性28例,阴性37例。对临床病理特征、CT影像学特征和IQon光谱CT定量参数的单因素分析显示,TNM分期(P<0.001)及肿瘤密度(P=0.025)在PD-L1阳性和阴性组间的差异有统计学意义,logistic回归分析结果显示TNM分期为独立预测因子,其训练集AUC为0.649,测试集AUC为0.611;经特征筛选后,最终12、6、8和6个影像组学特征分别构建基于CCI、IM、VNC图像及三者联合的影像组学模型。训练集中,CCI、IM、VNC图像和三者联合模型的AUC分别为0.729、0.692、0.699和0.728,测试集AUC则分别为0.639、0.662、0.662和0.717。将TNM分期与影像组学联合模型的Radscore相结合,建立临床-影像联合模型并绘制相应的诺模图,该模型训练集AUC为0.752,测试集AUC为0.725。Delong检验显示临床-影像联合模型在训练集(P=0.002)和测试集(P=0.037)的AUC显著高于临床模型,而与影像组学联合模型AUC无显著差异(P>0.05)。结论:基于双层探测器光谱CT的CCI、IM及VNC图像建立的影像组学模型在预测PD-L1表达方面具有一定的价值。结合Radscore与临床独立预测因子的临床-影像联合模型对PD-L1表达的预测价值较临床模型更高。根据该模型建立的诺模图可对NSCLC患者PD-L1表达实现个体化预测,能为临床制订免疫治疗的个体化方案提供参考。更多还原

【Abstract】 Object:To explore the predictive value of the dual-layer spectral detector CT-based radiomics features in PD-L1 expression of non-small cell lung cancer（NSCLC）.Method:Pathologically confirmed NSCLC patients,who underwent contrast-enhanced chest scan by dual-layer spectral detector CT（IQon spectral CT）before surgery from September 2019 to March 2021 in Tianjin Medical University Cancer Institute and Hospital,were enrolled prospectively.According to the inclusion and exclusion criteria,217 patients were finally included.Immunohistochemical method was used to detect the expression of PD-L1 in the surgically resected specimens of patients.The level of PD-L1 expression was expressed by tumor proportion score（TPS）and the patients were divided into PD-L1 positive groups and negative groups with a threshold of 1%.Patients were then randomly divided into a training set and a testing set in a 7:3 ratio.Clinicopathological characteristics of the patients were collected,including gender,age,smoking history,TNM stages and pathological types.CT imaging features were evaluated,including tumor location,size,density,shape,lobulation,spiculation,pleural retraction,vessel convergence sign,air bronchogram,and cavitation or bubblelike lucency.The quantitative parameters of IQon spectral CT were measured,including iodine concentration（IC）,normalized iodine concentration（NIC）,slope rate of the HU curve（λHU）and effective atomic number(Z_eff).120k Vp polyenergetic contrast-enhanced conventional images（CCI）,iodine map（IM）,and virtual non-contrast（VNC）images were saved in DICOM format from the PACS system and the IQon spectral CT post-processing workstation,respectively.Tumors were delineated around the outline slice by slice on CCI using ITK-SNAP software,and the three-dimensional region of interest（ROI）was finally obtained and copied to IM and VNC images.Python software was used to extract 1334 features from the ROI of the three sets of images,respectively.These features consist of first-order,second-order,shape,and transformed features.Feature screening and modeling were performed in four steps.The first step was to filter the features with interclass correlation coefficient（ICC）over 0.8 between the two observers.The second step was to standardize radiomics features by the Z-score standardization method.The third step includes two procedures.The first procedure was to select the features associated with PD-L1 expression（P<0.05）by independent sample t-test or Mann-Whitney U test.The second procedure was to evaluate the correlation of features by spearman or pearson correlation coefficients,and screen out the features with correlation coefficient above0.8.The fourth step was to use the LASSO-logistic method to establish a model for predicting the PD-L1 expression and calculate radiomics score（Radscore）.Four radiomics models based on CCI,IM,VNC images,and the combination of the three images were established.The differences in clinicopathological characteristics,CT imaging features,and spectral CT quantitative parameters between PD-L1 positive and negative groups were analyzed by univariate analysis.Multivariate logistic regression analysis was then conducted to select independent predictors.The independent predictors of PD-L1 were combined with Radscore of the combined radiomics model to establish a clinical-radiomics combined model and a nomogram for predicting PD-L1 expression.Feature screening and modeling were performed in the training set,and model diagnostic efficacy was tested in the testing set.Model diagnostic efficacy were evaluated by ROC curve.The comparison of AUC between models was performed using the Delong test.Results:There were 152 patients in the training set,of which 67 were PD-L1 positive and 85 were PD-L1 negative.65 patients were in the testing set,of which 28 were PD-L1 positive and 37 were PD-L1 negative.Significant differences were found in TNM stage（P<0.001）and tumor density（P=0.025）between PD-L1 positive and negative groups.Logistic regression analysis indicated that TNM stage was an independent predictor.For this model,the AUC in the training and testing set were 0.649 and 0.611,respectively.After feature screening,the remaining 12,6,8,and 6 radiomics features were used to construct radiomics models based on CCI,IM,VNC images,and the combination of the three sets of images respectively.In the training set,the AUC of models based on CCI,IM,VNC,and the combined images were 0.729,0.692,0.699,and 0.728,respectively,while AUC in the corresponding testing set were 0.639,0.662,0.662,and 0.717,respectively.The TNM stage was combined with the radscore of the combined radiomics model to establish a clinical-radiomics combined model.For this model,the AUC in training and testing set were 0.752 and 0.725,respectively.Delong test showed that the AUC of the clinical-radiomics combined model in the training set（P=0.002）and testing set（P=0.037）were superior to that of clinical model and not superior to radiomics model based on the combination of the three sets of images（P>0.05）.Conclusion:The radiomics model based on CCI,IM,and VNC images derived from dual-layer spectral detector CT may have potential for predicting PD-L1 expression.The clinical-radiomics combined model combined with radscore and independent clinical predictors has higher predictive value than the clinical model alone.The nomogram based on this model can visualize individualized prediction of PD-L1expression in NSCLC patients and be helpful in the planning of personalized immunotherapy.更多还原