基于^(18)F-FDG PET影像组学预测非小细胞肺癌病理亚型

Histological subtypes classification of non-small cell lung cancers using ^(18)F-FDG PET-based radiomics

目的探讨^(18)F-脱氧葡萄糖(FDG)PET影像组学特征在非小细胞肺癌(NSCLC)病理分型之间的差异,提高鉴别病理分型的能力。方法回顾分析2018年1月至2019年12月于华中科技大学同济医学院附属同济医院行^(18)F-FDG PET/CT的182例NSCLC患者[男109例,女73例;年龄(59.0±8.3)岁]。所有患者均经病理证实为肺腺癌或肺鳞状细胞癌(简称鳞癌)。使用简单随机抽样法将患者分为训练集(n=91)和验证集(n=91)。使用Python平台从PET图像中提取1132个影像组学特征。采用最大相关性最小冗余算法和最小绝对收缩和选择算子选择最优特征,并构建影像组学标签评分,使用Mann-Whitney U检验比较组间评分差异。用多因素logistic回归筛选病理亚型影响因素,基于临床变量和影像组学标签构建复合模型,并通过受试者工作特征(ROC)曲线评价模型预测能力,使用Delong检验模型曲线下面积(AUC)差异。结果4个影像组学特征[HHL_一阶最大值(first order_maximum)、LHL_一阶熵(first order_entropy)、HHH_灰度相关矩阵-大相关高灰度级强调(GLDM_LDHGLE)、HHL_GLDM_LDHGLE(H/L代表高或低通函数滤波处理)]被选择用于组学标签的构建。腺癌比鳞癌影像组学标签评分低[训练集:-1.30(-1.70,-1.04)与-0.60(-1.11,0.20),z=-4.61,P<0.001;验证集:-1.31(-1.66,-0.96)与-0.73(-1.02,-0.24),z=-4.76,P<0.001]。影像组学标签评分AUC[训练集:0.815(95%CI:0.723~0.906),验证集:0.813(95%CI:0.726~0.901)]高于临床变量AUC[吸烟史具有最优预测效能,训练集:0.721(95%CI:0.617~0.810),验证集:0.726(95%CI:0.623~0.814)],但差异无统计学意义(z值:1.319、1.324,均P>0.05)。综合临床变量(吸烟史)和组学标签构建的复合模型对病理亚型具有良好的辨别能力[训练集AUC=0.862(95%CI:0.785~0.940)、灵敏度88.00%(22/25)、特异性72.73%(48/66);验证集AUC=0.854(95%CI:0.776~0.933)、灵敏度75.00%(21/28)、特异性84.13%(53/63)],AUC高于临床变量(z值:3.257、3.872,均P<0.01)。结论由吸烟史、^(18)F-FDG PET影像组学标签形成的临床模型可为术前个体化预测NSCLC亚型提供非侵袭性、可重复的方法。

Objective To distinguish lung adenocarcinoma(ADC)from squamous cell carcinoma(SCC)using ^(18)F-fluorodeoxyglucose(FDG)PET-based radiomic features.Methods A retrospective analysis was performed in 182 patients(109 males,73 females,age(59.0±8.3)years)with non-small cell lung cancer(NSCLC)who underwent ^(18)F-FDG PET/CT scan between January 2018 and December 2019 in Tongji Hospital,Tongji Medical College,Huazhong University of Science and Technology.All patients had been diagnosed pathologically with lung ADC or SCC.The patients were divided into a training set(n=91)and a validation set(n=91)using simple random sampling method.Radiomic features were extracted from the PET images of segmented tumors using the Python package.The minimum redundancy maximum relevance feature selection algorithm and least absolute shrinkage and selection operator were employed to select informative and non-redundant features,and a radiomics signature score(rad-score)was developed.Differences of rad-score between groups were compared by Mann-Whitney U test.Multivariate logistic regression was applied to select the important factors.A combined model was constructed based on the clinical variable and radiomics signature.The predictive performance of models was analyzed and compared using receiver operating characteristic(ROC)curves and Delong test.Results Four radiomic features,namely HHL_first order_maximum,LHL_first order_entropy,HHH_gray level dependence matrix_large dependence high gray level emphasis(GLDM_LDHGLE),HHL_GLDM_LDHGLE(H/L represent the high/low pass filter)were selected to build the rad-score.The rad-score showed a significant ability to discriminate between different histological subtypes in the two sets(training set:-1.30(-1.70,-1.04)vs-0.60(-1.11,0.20),z=-4.61,P<0.001);validation set:-1.31(-1.66,-0.96)vs-0.73(-1.02,-0.24),z=-4.76,P<0.001).The area under the curve(AUC)of the rad-score were equal to 0.815(95%CI:0.723-0.906)in the training set,and 0.813(95%CI:0.726-0.901)in the validation set,respectively,which were larger than those of the clinical variables(smoking had the best prediction performance,training set:0.721(95%CI:0.617-0.810),validation set:0.726(95%CI:0.623-0.814)),however,the difference was not significant(z values:1.319,1.324,both P>0.05).When the clinical variable(smoking)and radiomics signature were combined,the complex model showed a better performance in the classification of histological subtypes,with the AUC increased to 0.862(95%CI:0.785-0.940;sensitivity:88.00%(22/25),specificity:72.73%(48/66))in the training set and 0.854(95%CI:0.776-0.933;sensitivity:75.00%(21/28),specificity:84.13%(53/63))in the validation set.The AUC values were significantly higher than those of the clinical variable(smoking;training set:z=3.257,P<0.001;validation set:z=3.872,P<0.001).Conclusion Individualized diagnosis model incorporating with smoking and radiomics signature can help differentiate lung cancer subtypes in a non-invasive,repeatable modality.