基于^(18)F-FDG PET/CT影像组学的机器学习模型预测非小细胞肺癌病理分化程度的价值  

The value of machine learning models based on ^(18)F-FDG PET/CT radiomics for predicting the degree of tumour differentiation in patients with non-small cell lung cancer

作　　者：于军 李洋 杨雪 毕晓峰 任东栋 任春玲 黄磊 YU Jun;LI Yang;YANG Xue;BI Xiaofeng;REN Dongdong;REN Chunling;HUANG Lei(Department of Nuclear Medicine,Ningbo Mingzhou Hospital,Ningbo 315100,China)

机构地区：[1]宁波明州医院核医学科,浙江宁波315100

出　　处：《温州医科大学学报》2024年第9期709-717,共9页Journal of Wenzhou Medical University

摘　　要：目的:探讨基于^(18)F-FDG PET/CT影像组学的不同机器学习模型预测非小细胞肺癌(NSCLC)病理分化程度的价值。方法:回顾性纳入2019年1月至2023年8月于宁波明州医院行18F-FDG PET/CT检查并行根治性手术的325例NSCLC患者(男191例,女134例,年龄40~85岁),其中非低分化(高、中分化)NSCLC患者157例,低分化NSCLC患者168例。采用随机法,按7:3比例将患者分为训练集(227例)与验证集(98例)。使用LIFEx 7.4.3软件提取原发肿瘤病灶PET/CT影像组学特征,采用最小绝对收缩与选择算子(LASSO)算法及10倍交叉验证进行特征筛选,构建7种机器学习模型:决策树(DT)模型、随机森林(RF)模型、K最近邻(KNN)模型、朴素贝叶斯(NB)模型、极限梯度提升(XGBoost)模型、支持向量机(SVM)模型及Logistic回归(LR)模型。采用ROC曲线分析评估各种模型的预测效能。结果:从^(18)F-FDG PET/CT图像中共提取250个影像组学特征,经LASSO算法及10倍交叉验证最终筛选出10个最佳组学特征,在构建的7种机器学习模型中,DT、RF、XGBoost模型在训练集中的AUC分别为0.858、0.951、0.936,在验证集中分别降至为0.594、0.694、0.668,存在明显的过拟合现象。KNN、NB、SVM、LR模型在训练集中的AUC分别为0.773、0.759、0.801、0.761,在验证集中的AUC分别为0.680、0.668、0.726、0.688,具有较强的泛化能力及稳定性。Delong检验显示,在训练集及验证集中,KNN、NB、SVM、LR模型两两间AUC值比较差异均无统计学意义(均P>0.05)。结论:基于^(18)F-FDG PET/CT影像组学的机器学习模型可有效地预测NSCLC患者病理分化程度,KNN、NB、SVM、LR模型在训练集和验证集中均具有较高的AUC值,可辅助临床决策及制定个性化治疗方案。Objective:To investigate the value of different machine learning models based on ^(18)F-FDG PET/CT radiomics for predicting the degree of tumour differentiation in non-small cell lung cancer(NSCLC).Methods:Three hundred and twenty-five patients with NSCLC(191 males,134 females,aged 40-85 years)who underwent ^(18)F-FDG PET/CT followed by radical surgery from January 2019 to August 2023 were retrospectively enrolled,including 157 patients with non-poorly differentiated and 168 patients with poorly differentiated.Patients were randomly divided into a training cohort(227 cases)and a validation cohort(98 cases)in a 7:3 ratio.LIFEx 7.4.3 software was used to extract the PET/CT radiomics features,and the least absolute shrinkage and selection operator(LASSO)method and 10-fold cross-validation were used for feature screening.Seven machine learning models,namely decision tree(DT),random forest(RF),K-nearest neighbour(KNN),naive bayesian(NB),extreme gradient boosting(XGBoost),support vector machine(SVM),logistic regression(LR)models,were constructed based on the selected optimal feature subsets.The ROC curve analysis was used to assess the predictive ability of various models.Results:A total of 250 radiomics features were extracted from PET/CT images,and 10 radiomics features were finally screened by the LASSO algorithm and 10-fold cross-validation,including 5 PET features and 5 CT features.Among the seven machine learning models constructed,the AUCs of the DT,RF,and XGBoost models in the training cohort were 0.858,0.951,and 0.936,respectively,and decreased to 0.594,0.694,and 0.668 in the validation cohort,with obvious overfitting.The AUCs of KNN,NB,SVM,LR models in the training cohort were 0.773,0.759,0.801,0.761,and in the validation cohort were 0.680,0.668,0.726,0.688,respectively,which have strong generalisation ability and stability.Conclusion:The machine learning models based on ^(18)F-FDG PET/CT radiomics can effectively predict the degree of tumour differentiation in patients with NSCLC,and the KNN,NB,SVM and LR mo

关 键 词：非小细胞肺癌 肿瘤分化 正电子发射断层显像术 体层摄影术 X线计算机 影像组学 机器学习 

分 类 号：R734.2[医药卫生—肿瘤] R445.6[医药卫生—临床医学]