基于CT检查影像组学预测模型对结直肠癌肝转移灶组织病理学生长方式的预测价值

Predictive value of radiomics model based on CT for histopathological growth patterns in tissues of colorectal cancer liver metastases

目的探讨基于CT检查影像组学预测模型对结直肠癌肝转移(CRCLM)灶组织病理学生长方式(HGP)的预测价值.方法采用回顾性队列研究方法.收集2017年5月至2022年5月温州市中心医院收治的100例CRCLM患者临床病理资料;男56例,女44例;年龄为(54±13)岁.患者术前均行腹部增强CT检查,均行根治性切除术.100例患者通过随机数字表法按7∶3分为训练集70例和验证集30例.训练集用于构建预测模型,验证集用于验证预测模型.正态分布的计量资料以x±s表示,组间比较采用t检验;偏态分布的计量资料以M(Q_(1),Q_(3))表示,组间比较采用Mann-Whitney U检验.计数资料以绝对数表示,组间比较采用χ^(2)检验.单因素分析采用对应的统计学方法.多因素分析采用Logistic回归分析.绘制受试者工作特征曲线并计算曲线下面积(AUC)、准确度、灵敏度、特异度.结果(1)影像组学特征提取及筛选.提取100例CRCLM患者1635个影像组学特征,筛选组内相关系数>0.8的特征,经最大相关最小冗余算法及最小绝对收缩和选择算子方法回归分析降维,最终筛选6个影像组学特征(wavelet-LHL_firstorder_Kurtosis、lbp-2D_firstorder_Median、original_shape_Spgericty、original_shape_VolxelVolume、wavelet-HLH_glrlm_GLNUN、wavelet-HLL_glcm_Imc2).(2)影响训练集CRCLM患者HGP类型的因素.多因素分析结果显示:术前癌胚抗原和术前CA19-9是训练集CRCLM患者HGP类型的独立影响因素(优势比=16.83,3.26,95%可信区间为1.04~258.60,1.07~19.32,P<0.05).(3)HGP预测模型的构建与评价.基于多因素分析结果、影像组学特征,通过随机森林机器学习方法构建临床-影像组学联合预测模型.训练集中联合预测模型的AUC为0.98(95%可信区间为0.96~1.00)、准确度为0.91、灵敏度为0.89、特异度为0.90;验证集中联合预测模型的AUC为0.84(95%可信区间为0.63~1.00)、准确度为0.90、灵敏度为0.93、特异度为0.88.结论基于CT检查影像组学结合临床病理资料多因素分析结果共同构建的临床-影像组学联合预测模型对CRCLM的HGP具体类型有预测作用,且效能较好.

Objective To investigate the predictive value of radiomics predictive model based on computed tomography(CT)for histopathological growth patterns(HGP)in tissues of colo-rectal cancer liver metastases(CRCLM).Methods The retrospective cohort study was conducted.The clinicopathological data of 100 CRCLM patients who were admitted to Wenzhou Central Hospital from May 2017 to May 2022 were collected.There were 56 males and 44 females,aged(54±13)years.All patients underwent preoperative enhanced abdominal CT and radical resection.A total of 100 CRCLM patients were randomly divided into 70 cases in the training set and 30 cases in the valida-tion set according to the ratio of 7∶3 based on random number table.The training set was used to construct predictive model,and validation set was used to validate predictive model.Measurement data with normal distribution were represented as Mean±SD,and comparison between groups was analyzed using the t test.Measurement data with skewed distribution were expressed as M(Q_(1),Q_(3)),and comparison between groups was conducted using the Mann-Whitney U test.Count data were represented as absolute numbers,and comparison between groups was conducted using the chi-square test.Univariate analysis was performed using statistical methods appropriate to the data type.Multivariate analysis was conducted using the Logistic regression model.The receiver operating characteristic curve was drawn,and the area under curve,accuracy,sensitivity and specificity were calculated.Results(1)Radiomics featuers extraction and selection.A total of 1635 radiomics features of 100 CRCLM patients were extracted to select features with intra-group correlation coefficient>0.8.After dimension reduction of features by using the maximal relevance and minimal redundancy,the Least Absolute Shrinkage and Selection Operator regression analysis,6 radiomics features(wavelet-LHL_firstorder_Kurtosis,lbp-2D_firstorder_Median,original_shape_Spgericty,original_shape_VolxelVolume,wavelet-HLH_glrlm_GLNUN,wavelet-HLL_glcm_Imc2)were finally screened out.(2)Analysis of influencing factors for HGP in tissues of CRCLM patients of the training set.Results of multivariate analysis showed that preoperative carcinoembryonic antigen and CA19-9 were indepen-dent influencing factors for HGP in tissues of CRCLM patients of the training set(odds ratio=16.83,3.26,95%confidence interval as 1.04-258.60,1.07-19.32,P<0.05).(3)Construction and evaluation of predictive model for HGP.The clinical-radiomics combined predictive model was constructed based on the results of multivariate analysis and rediomics features using the random forest machine learning method.The AUC of clinical-radiomics combined predictive model in the training set was 0.98(95%confidence interval as 0.96-1.00),with the accuracy,sensitivity and specificity as 0.91,0.89 and 0.90.The AUC of clinical-radiomics combined predictive model in the validation set was 0.84(95%confidence interval as 0.63-1.00),with the accuracy,sensitivity and specificity as 0.90,0.93 and 0.88.Conclusion The clinical-radiomics combined predictive model based on CT radio-mics and multivariate analysis of clinicopathological data has predictive value for HGP in CRCLM,with good performance.