增强CT检查预测肝细胞癌微血管侵犯的影像基因组学研究

Radiogenomics of enhanced cT imaging to predict microvascular invasion in hepatocellular carcinoma

目的构建基于术前增强CT检查的联合影像组学模型,预测肝细胞癌微血管侵犯(MVI)状态,对影像组学模型进行生物学解释.方法采用回顾性队列研究方法.收集癌症基因组图谱数据库建库至2023年1月纳入的424例肝细胞癌患者的mRNA数据,癌症图像档案馆数据库建库至2023年1月纳入的39例肝细胞癌患者和甘肃省人民医院2020年1月至2023年1月收治53例肝细胞癌患者的临床病理资料.92例肝细胞癌患者通过随机数字表法按7∶3分为训练集64例和测试集28例.分析动脉期及门静脉期CT检查图像及临床资料.使用3Dslicer软件(5.0.3版本)进行动脉期和门静脉期图像配准和三维感兴趣区勾画.使用开源软件FAE(0.5.5版本)对原始图像进行预处理并提取特征.通过最小绝对收缩和选择算子等方法筛选特征,构建影像组学模型并计算影像组学评分(R-score),通过Logistic回归整合临床参数、影像学特征及R-score构建列线图.通过加权基因共表达网络分析和相关性分析获取影像组学模型相关的基因模块并进行富集分析.观察指标:(1)不同MVI性质患者的临床特征比较.(2)MVI风险模型的建立.(3)MVI风险模型的评估.(4)基因模块聚类.(5)特征相关基因模块功能富集.正态分布的计量资料以(x)±s表示,组间比较采用独立样本t检验,偏态分布的计量资料以M(范围)表示,组间比较采用Mann-Whitney U检验,计数资料比较采用χ^(2)检验.采用组内和组间相关系数(ICC)评估影像组学特征提取的观察者间的一致性.ICC>0.75表示特征提取的一致性良好.单因素和多因素分析采用Logistic回归模型.绘制受试者工作特征曲线,以曲线下面积(AUC)、决策曲线、校准曲线评估模型的诊断效能及临床实用性.结果(1)不同MVI性质患者的临床特征比较.92例肝细胞癌患者中,MVI阳性47例,MVI阴性45例,两者肝炎、肿瘤长径、瘤周增强、瘤内动脉、假包膜及瘤周不光滑比较,差异均有统计学意义(χ^(2)=5.308,9.977,47.370,32.368,21.105,31.711,P<0.05).(2)MVI风险模型的建立.在动脉期及门静脉期的瘤内和瘤周分别提取了1781个特征,经过特征降维后,从动脉期及门静脉期中确定8个影像组学特征构建联合模型.多因素分析结果显示:瘤周增强、瘤内动脉、假包膜、瘤周不光滑及R-score是肝细胞癌患者MVI的独立危险因素[风险比=0.049,0.017,0.017,0.021,2.539,95%可信区间(CI)为0.005~0.446,0.001~0.435,0.001~0.518,0.001~0.473,1.220~3.283].纳入瘤周增强、瘤内动脉、假包膜、瘤周不光滑及R-score构建列线图模型.(3)MVI风险模型的评估.R-score在训练集和测试集中AUC分别为0.923(95%CI为0.887~0.944)和0.918(95%CI为0.894~0.945);联合R-score及影像学特征构建的列线图在训练集和测试集中AUC分别为0.973(95%CI为0.954~0.988)和0.962(95%CI为0.942~0.987).决策曲线显示:列线图的临床效益优于R-score.校准曲线显示:列线图和R-score预测状态与实际观察结果间一致性良好.(4)基因模块聚类.经加权基因共表达网络分析后获取8个基因模块.(5)特征相关基因模块功能富集.4个基因模块与影像组学特征显著相关.预测MVI的影像组学特征可能与细胞周期、中性粒细胞外陷阱形成及PPAR信号通路有关.结论基于术前增强CT检查的联合影像组学模型可以预测肝细胞癌MVI状态.通过获取影像组学特征相关的mRNA基因表达谱,为影像组学模型提供了生物学解释.

Objective To construct a combined radiomics model based on preoperative enhanced computed tomography(CT)examination for predicting microvascular invasion(MVI)in hepatocellular carcinoma(HCC),and provide biological explanations for the radiomics model.Methods The retrospective cohort study was conducted.The messenger RNA(mRNA)of 424 HCC patients,the clinicopathological data of 39 HCC patients entered into the Cancer Genome Atlas database from its establishment until January 2023,and the clinicopathological data of 53 HCC patients who were admitted to the Gansu Provincial People's Hospital from January 2020 to January 2023 were collected.The 92 HcC patients were randomly divided into a training dataset of 64 cases and a test dataset of 28 cases with a ratio of 7:3 based on a random number table method.The CT images of patients in the arterial phase and portal venous phase as well as the corresponding clinical data were analyzed.The 3Dslicer software(version 5.0.3)was used to register the CT images in the arterial phase and portal venous phase and delineate the three-dimensional regions of interest.The original images were preprocessed and the corresponding features were extracted by the open-source software FAE(version 0.5.5).After selecting features using the Least Absolute Shrinkage and Selection Operator,the radiomics model was constructed and the radiomics score(R-score)was calculated.The nomogram was constructed by integrating clinical parameters,imaging features and R-score based on Logistic regression.The gene modules related to radiomics model were obtained and subjected to enrichment analysis by conducting weighted gene co-expression network analysis and correlation analysis.Observation indicators:(1)comparison of clinical characteristics of patients with different MVI properties;(2)establishment of MVI risk model;(3)evaluation of MVI risk model;(4)clustering of gene modules;(5)functional enrichment of feature-correlated gene modules.Measurement data with normal distribution were represented as Mean+SD,and comparison between groups was conducted using the independent sample t test.Measurement data with skewed distribution were represented as M(range),and comparison between groups was conducted using the Mann-Whitney U test.Comparison of count data was conducted using the chi-square test.The intra-/interclass correlation coefficient(ICC)was used to assess the inter-observer consistency of radiomics feature extracted by different observers.ICC>0.75 indicated a good consistency in feature extraction.The Logistic regression model was used for univariate and multivariate analyses.The receiver operating characteristic curve was drawn,and the area under curve(AUC),the decision curve and the calibration curve were used to evaluate the diagnostic efficacy and clinical practicality of the model.Results(1)Comparison of clinical characteristics of patients with different MVI properties.Of 92 HCC patients,there were 47 cases with MVI-positive and 45 cases with MVI-negative,and there were significant differences in hepatitis,tumor diameter,peritumoral enhancement,intratumoral arteries,pseudocapsule and smoothness of tumor margin between them(χ^(2)=5.308,9.977,47.370,32.368,21.105,31.711,P<0.05).(2)Establishment of MVI risk model.A total of 1781 features were extracted from arterial and portal venous phases of the intratumoral and peritumoral regions.After feature dimension reduction,8 radiomics features were selected from arterial and portal venous phases to construct the combined model.Results of multivariate analysis showed that peritumoral enhancement,intratumoral arteries,pseudocapsule,smoothness of tumor margins,and R-score were independent risk factors for MVI in patients with HCC[hazard ratio=0.049,0.017,0.017,0.021,2.539,95%confidence interval(c)as 0.005-0.446,0.001-0.435,0.001-0.518,0.001-0.473,1.220-5.283,P<0.05].A nomogram model was constructed incorporating peritumoral enhancement,intratumoral arteries,pseudocapsule,smoothness of tumor margins,and R-score.(3)Evaluation of the MVI risk model.The AUC of radiomics model was 0.923(95%CI as 0.887-0.944)and 0.918(95%CI as 0.894-0.945)in the training dataset and test dataset,respectively.The AUC of nomogram model,incorporating both the R-score and radiomics features,was 0.973(95%Cl as 0.954-0.988)and 0.962(95%Cl as 0.942-0.987)in the training dataset and test dataset,respectively.Results of decision curve showed that the nomogram had better clinical utility compared to the R-score.Results of calibration curve showed good consistency between the actual observed outcomes and the nomogram or the R-score.(4)Clustering of gene module.Results of weighted gene co-expression network analysis showed that 8 gene modules were obtained.(5)Functional enrichment of feature-related gene modules.Results of correlation analysis showed 4 gene modules were significantly associated with radiomics features.The radiomics features predicting of MVI may be related to pathways such as the cell cycle,neutrophil extracellular trap formation,and PPAR signaling pathway.Conclusions The combined radiomics model based on preoperative enhanced CT imaging can predict the MVI status of HCC.By obtaining mRNA gene expression profiles associated with radiomics features,a biological interpretation of the radiomics model is provided.