基于^(18)F-FDG PET/CT影像组学分析不同机器学习模型预测非小细胞肺癌隐匿性淋巴结转移的价值

Analysis on different ^(18)F-FDG PET/CT radiomics-based machine learning models for predicting occult lymph node metasta-sis in non-small cell lung cancer

目的探讨基于治疗前18氟-脱氧葡萄糖(^(18)F-FDG)正电子发射断层显像术(PET)/CT影像组学预测非小细胞肺癌(NSCLC)患者隐匿性淋巴结转移(OLM)的价值,分析不同机器学习模型对预测结果的影响。方法回顾性选取2019年1月至2023年5月于宁波明州医院行^(18)F-FDG PET/CT检查并行根治性手术及系统性淋巴结清扫的NSCLC患者324例(男186例,女138例,年龄36~85岁),其中OLM阴性258例,阳性66例。采用随机数字表法按7:3比例将患者分为训练集(226例)与验证集(98例)。使用LIFEx 7.4.3软件提取病灶PET/CT影像组学特征,采用最小绝对收缩与选择算子(LASSO)算法进行特征筛选,构建3种机器学习模型:逻辑回归(LR)模型、支持向量机(SVM)模型、随机森林(RF)模型。采用ROC曲线分析评估各种模型的预测效能,并采用决策曲线(DCA)分析各种模型的临床价值。结果从PET/CT图像中共提取出250个影像组学特征,经LASSO算法最终筛选出8个组学特征,包括4个PET特征[直方图(HISTO)_均匀性(Uniformity)、灰度共生矩阵(GL-CM)_差熵(DE)、灰度游程长度矩阵(GLRLM)_短行程低灰度强调(SRLGLE)、灰度区域大小矩阵(GLSZM)_小区域低灰度强调(SZLGLE)],4个CT特征[形态(MORPH)_质量中心偏移(CMS)、HISTO_四分位离散系数(QCD)、HISTO_最大直方图梯度(MHG)、GLSZM_大区域强调(LZE)]。在构建的3种机器学习模型中,以SVM模型预测效能最优,其在训练集及验证集中的AUC分别为0.846、0.849;LR模型在训练集与验证集中的AUC分别为0.696、0.711;RF模型在训练集与验证集中的AUC分别为0.943、0.568,存在明显的过拟合现象。DCA分析显示,SVM模型及LR模型均具有较好的净获益与临床价值。结论基于^(18)F-FDG PET/CT影像组学分析可有效预测NSCLC患者是否存在OLM,SVM模型预测性能最佳,可辅助临床决策及制定个性化治疗方案。

Objective To investigate the value of pre-treatment ^(18)F-FDG PET/CT radiomic analysis in predicting occult lymph node metastasis(OLM)of non-small cell lung cancer(NSCLC),and to evaluate the influence of different machine learning models on the predictive outcomes.Methods Three hundred and twenty-four patients with NSCLC(186 males,138 females,aged 36-85 years)who underwent ^(18)F-FDG PET/CT examination followed by radical surgery and systematic lymph node dissection at Ningbo Mingzhou Hospital from January 2019 to May 2023 were retrospectively enrolled.Among them,258 cases were OLM-negative and 66 cases were OLM-positive.Patients were randomly divided into a training set(226 cases)and a validation set(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 was used for feature screening.Three machine learning models,namely logistic regression(LR),support vector machine(SVM),and random forest(RF)models,were constructed based on the selected optimal feature subsets.The ROC curve analysis was used to assess the predictive ability of the models,and decision curve analysis(DCA)was used to analyze their clinical values.Results A total of 250 radiomics features were extracted from PET/CT images,and eight were finally screened out by the LASSO algorithm,including four PET features[histogram(HISTO)_Uniformity,grey level co-occurrence matrix(GLCM)_Difference Entropy(DE),grey level run length matrix(GLRLM)_Short Run Low Grey Level Emphasis(SRLGLE),and grey level size zone matrix(GLSZM)_Small Zone Low Grey Level Emphasis(SZLGLE)],and four CT features[morphological(MORPH)_Centre Of Mass Shift(CMS),HISTO_Quartile Coefficient of Dispersion(QCD),HISTO_Maximum Histogram Gradient(MHG),and GLSZM_Large Zone Emphasis(LZE)].Of the three constructed machine learning models,the SVM model demonstrated the most effective predictive performance,with AUCs of 0.846 and 0.849 in the training and validation sets,respectively.The LR model had AUCs of 0.696 and 0.711 in the training and validation sets,respectively;while the RF model obtained AUCs of 0.943 and 0.568 in the training and validation sets,respectively,and an obvious over-fitting phenomenon was observed.DCA demonstrated that both the SVM model and the LR model had superior clinical value and net benefit.Conclusion ^(18)F-FDG PET/CT radiomics analysis can effectively predict the presence of OLM in patients with NSCLC,and the SVM model demonstrated the best predictive performance,which may help in clinical decision-making and the formulation of individualized treatment plans.