基于多序列MRI影像组学评分及临床参数预测肝细胞癌微血管浸润的nomogram模型研究  

A nomogram model for predicting microvascular invasion of hepatocellular carcinoma based on multi-sequences MRI radiomics score and clinical-pathology-imaging parameters

作　　者：刘小芳 汪清华 杨洪安 肖琼 廖建 高阳 谭永明[2] LIU Xiao-fang;WANG Qing-hua;YANG Hong-an(Department of Radiology,the First Affiliated Hospital of Nanchang University,Nanchang 330006,China)

机构地区：[1]新余市人民医院影像中心,江西新余336500 [2]南昌大学第一附属医院影像科,江西南昌330006

出　　处：《放射学实践》2023年第8期1018-1025,共8页Radiologic Practice

基　　金：江西省医学影像临床研究中心项目(20223BCG74001);江西省研究生创新专项资金项目(YC2023-B078)。

摘　　要：目的:基于肝细胞癌(HCC)患者的临床资料及多模态肝脏影像组学分析建立机器学习模型,探讨此模型术前预测HCC微血管浸润(MVI)的价值。方法:回顾性分析2020年3月-2021年9月在本院经病理证实为原发性HCC的130例患者的术前肝脏MRI及临床资料。基于病理检查结果,将患者分为MVI阳性组及MVI阴性组。记录患者的各项术前临床资料。所有患者术前行MRI检查,检查序列包括T 2WI、DWI和ADC以及Gd-EOB-DTPA对比增强动脉期、门脉期、延迟期和肝胆期T_(1)WI共7个序列。由放射科医师评估肿瘤的常规影像特征。自7个序列的图像上分别提取影像组学特征并进行降维,然后采用线性支持向量机(SVM)方法构建预测MVI的预测模型。再将所有序列图像提取的特征整合,经降维分析后最终筛选出6个最佳组学特征并采用线性SVM方法构建多序列联合组学模型,然后基于此多序列联合组学模型计算每例患者的放射组学评分(Radscore)作为后续建模特征。最后共采用了5种机器学习算法对上述三类资料(即临床资料、常规影像特征、组学特征)中筛选出的特征进行综合模型的构建,包括线性的支持向量机(linear SVM)、带rbf核函数的支持向量机(rbf-SVM)、逻辑回归(LR)、随机森林(RF)和XGBoost(XGB)。采用ROC曲线及概率校准曲线验证并评估单一或联合模型预测MVI的效能,根据受试者工作特性(ROC)曲线下面积选择最优模型。结果:临床指标及常规影像学特征中甲胎蛋白浓度、动脉期瘤周增强、肿瘤最大直径、肿瘤边缘、肿瘤生长模式、瘤内出血以及静脉侵犯征象(RVI)是MVI的独立预测因子。在7种单序列及多序列联合组学模型中,以多序列联合模型的诊断效能为最佳(在训练集中的AUC=0.913,95%CI:0.822~1.000)。建立的5个机器学习综合模型中rbf-SVM模型的预测效能最好。相较于Radscore(测试集:AUC=0.879,95%CI=0.755~1.000)、临床病理(测�Objective:Established a radiomics machine learning model based on multimodal MRI and clinical data,and to analyze the preoperative prediction value of this model for microvascular invasion(MVI)of hepatocellular carcinoma(HCC).Methods:The preoperative liver MRI data and clinical information of 130 patients with pathologically confirmed HCC were retrospectively studied.The patients were divided into MVI-positive(MVI+)group and MVI-negative(MVI-)group based on postoperative pathology.The preoperative clinical indicators of patients were recorded,including AFP,ALB,ALP,ALT,AST,APTT,CA12-5,CA12-9,CB,CEA,FIB,GR,PT,TB andγ-GLU.All patients underwent preoperative MRI examination,7 sequences were scanned,including T 2WI,DWI,ADC,and Gd-EOB-DTPA contrast-enhanced T_(1)WI of four phases(arterial phase,portal phase,delayed phase,and hepatobiliary phase).The conventional imaging features of tumors on the 7 sequences were evaluated by a radiologist.After a series of dimensionality reduction analysis,six features were finally screened out after extracting image omics features from seven sets of images,and then a prediction model of preoperative microvascular invasion was established using linear support vector machine(li-near-SVM),support vector machine with rbf kernel function(rbf-SVM),logistic regression(LR),Random forest(RF)and XGBoost(XGB)algorithms.Based on the area under the receiver operating characteristic(ROC)curve,the model with the best performance was selected out,and the Radscore of each patients was calculated.Finally,establish Radscore,clinical pathology conventional imaging prediction models,and combined nomogram models,and draw column charts of combined nomogram mo-dels.Verify and evaluate the predictive MVI performance of single or joint models using ROC and probability calibration curves.Results:Among clinical indicators and conventional imaging features,Alpha-fetoprotein concentration,peritumor enhancement,maximum tumor diameter,smooth tumor margins,tumor growth pattern,presence of intratumor hemorrhage,and R

关 键 词：影像组学 磁共振成像 肝细胞癌 微血管浸润 机器学习 预测模型 

分 类 号：R445.2[医药卫生—影像医学与核医学] R735.7[医药卫生—诊断学]