CT影像组学预测胰腺导管腺癌患者术后无病生存期的价值
The value of CT radiomics in predicting postoperative disease-free survival in patients with pancreatic ductal adenocarcinoma
魏显飞 1陈基明 1吴莉莉 1何天洪 1谷士康 1刘美娜 1谢伶俐1
作者信息
- 1. 241001 安徽芜湖,皖南医学院弋矶山医院影像中心
- 折叠
摘要
目的:探讨基于平扫和三期增强CT的影像组学模型及临床-组学综合模型对胰腺导管腺癌(PDAC)患者术后无病生存期(DFS)的预测价值.方法:回顾性分析2013年12月-2021年6月在本院经术后病理证实的124例胰腺导管腺癌患者的病例资料.所有DFS患者术后随访时间大于3个月.采用随机分组法,按照7∶3的比例将患者分为训练集(n=87)和验证集(n=37).所有患者术前行腹部CT平扫及三期(动脉期、静脉期、延迟期)增强扫描.使用ITK-SNAP软件分别在四期CT图像上沿胰腺肿瘤边缘逐层勾画ROI并融合生成三维容积ROI(VOI),然后导入FAE软件中提取影像组学特征.采用单因素Cox回归分析及LASSO-Cox回归分析进行纹理特征的筛选,然后分别构建各期和多期联合(动脉期+静脉期+延迟期)影像组学模型并计算相应的影像组学标签得分.采用单因素和多因素Cox回归分析筛选临床特征和CT形态学特征并构建临床模型.采用多因素Cox回归分析结合临床模型变量及影像组学标签构建临床-组学综合模型并绘制其诺莫图.采用一致性指数(C-index)、时间依赖性(time-dependent)ROC曲线、校正曲线和决策曲线分析(DCA)对模型的诊断效能及临床效益进行评价.利用R语言计算临床-组学综合模型的最佳截断值,并据此将患者分为高风险组和低风险组,采用Kaplan-Meier法分析生存资料并进行log-rank检验.结果:基于平扫、动脉期、静脉期和延迟期及多期联合分别筛选得到5、16、4、12和17个组学特征,分别建立相应的组学模型并获得影像组学标签值.经log-rank检验,所有组学标签均与DFS具有相关性(P＜0.05),其中多期联合模型的预测效能最佳(训练集:C-index=0.786,6～24 个月 AUC=0.850～0.928;验证集:C-index=0.802,6～24 个月AUC=0.796～0.874);而临床模型的预测效能较低(训练集:C-index=0.635,6～24个月AUC=0.647～0.679;验证集:C-index=0.596,6～24个月AUC=0.545～0.656).临床-组学综合模型的预测效能(训练集:C-index=0.812,6～24 个月 AUC=0.883～0.958;验证集:C-index=0.796,6～24 个月AUC=0.813～0.894)明显优于临床模型;校准曲线显示临床-组学综合模型的拟合度好;DCA显示临床-组学综合模型的临床净收益优于临床模型.临床-组学综合模型的截断值为2.738.Kaplan-Meier生存分析显示在训练集和验证集中,高风险组患者的DFS明显短于低分风险组.结论:基于多期CT扫描的影像组学模型结合临床特征构建的临床-组学综合模型在预测胰腺导管腺癌患者术后DFS方面,相较于临床模型和影像组学模型具有更好的预测效能,有助于指导临床制订个体化的治疗策略和改善患者的预后.
Abstract
Objective:To investigate the value of CT radiomics and clinical-radiomics compre-hensive model based on pre-contrast and three phases contrast-enhanced CT in predicting postopera-tive disease-free survival(DFS)in patients with pancreatic ductal adenocarcinoma.Methods:A retro-spective analysis was performed for the data of 124 patients with pancreatic ductal adenocarcinoma confirmed by postoperative pathology in our hospital from December 2013 to June 2021.All patients with DFS were followed up more than three months.The patients were divided into training set(n=87)and validation set(n=37)according to the 7∶3 ratio using the randomization method.All patients underwent pre-contrast and three-phase(arterial phase,venous phase and delayed phase)contrast-en-hanced abdominal CT scan before surgery.ITK-SNAP software was used to delineate the regions of in-terest(ROI)along the edge of the lesions on each of the four phases CT images,and three-dimensio-nal fusion was performed to obtain volume ROI(VOI),and then they were imported into FAE soft-ware to extract the radiomics features of each stage.Univariate Cox regression and LASSO-Cox re-gression analysis were used to screen texture features,and radiomics labels were constructed and ra-diomics scores were calculated for each phase and three-phase.Univariate and multivariate Cox regres-sion was used to screen clinical features and CT features and construct clinical model.Multivariate Cox regression analysis was used to establish clinical-radiomics comprehensive model combining clinical data,CT morphology features and radiomics labels,and nomogram was drawn.Model performance and clinical benefit were evaluated by concordance index(C-index),time-dependent ROC,correction curve and decision curve analysis(DCA).The optimal cut-off of the clinical-radiomics comprehensive model was calculated using R Programming Language to divide patients into high-risk and low-risk groups,the Kaplan-Meier method was used to analyze survival data and perform log-rank test.Results:5,16,4,12 and 17 features were selected out based on the pre-contrast,arterial phase,venous phase,delayed phase and multi-phase,respectively.And then the corresponding radiomics models were established and values of radiomics labels were obtained.Verified by log-rank,all labels were associated with DFS(all P＜0.05),among them,the multiphase model showed the best performance(in training set:C-in-dex=0.786,6～24 months AUC=0.850～0.928;in validation set:C-index=0.802,6～24 months AUC=0.796～0.874).Clinical model prediction performance was low(in training set:C-index=0.635,6～24 months AUC=0.647～0.679;in validation set:C-index=0.596,6～24 months AUC=0.545～0.656).Predictive efficacy of the clinical-radiomics comprehensive model(in training set:C-index=0.812,6～24 months AUC=0.883～0.958;in validation set:C-index=0.796,6～24 months AUC=0.813～0.894)was significantly better than that of the clinical model.The calibration curve showed that the clinical-radiomics comprehensive model fits well,and the DCA showed that the clinical benefit of the clinical-radiomics comprehensive model was better than that of the clinical model.The cut-off of the clinical-radiomics comprehensive model was 2.738.Kaplan Meier survival analysis showed that the DFS of high-risk patients in the training set and validation set was significantly shorter than that of the low-risk group.Conclusion:The clinical-radiomics comprehensive model based on multi-phase CT radiomics labels combined with clinical features in predicting postoperative DFS in patients with pan-creatic ductal adenocarcinoma has better predictive performance than clinical models and radiomics models,this may help to guide clinical the formulation of individualized treatment strategy and im-prove the prognosis of patients.
关键词
胰腺肿瘤/导管腺癌/无病生存期/影像组学/预测模型/体层摄影术,X线计算机Key words
Pancreatic neoplasms/Ductal adenocarcinoma/Disease-free survival/Radiomics/Predictive model/Tomography,X-ray computed引用本文复制引用
出版年
2024
NETL
NSTL
万方数据