基于瘤内及瘤周区域计算机断层扫描放射组学特征联合临床因素机器学习模型预测肝细胞癌磷脂酰基醇蛋白聚糖3表达  

Machine learning model based on CT radiomics features of intratumoral and peritumoral regions combined with clinical factors for predicting glypican-3 expression in hepatocellular carcinoma

作　　者：朱梓维 吴军 郭杨 任琼渊 李冬宁 李泽雨 韩磊 李川 ZHU Zi-wei;WU Jun;GUO Yang;REN Qiong-yuan;LI Dong-ning;LI Ze-yu;HAN Lei;LI Chuan(China Medical University,General Hospital of Northern Theater Command Training Base for Graduate,Shenyang 110016,China;Department of Hepatobiliary Surgery;Department of Medical Service,General Hospital of Northern Theater Command,Shenyang 110016,China;Dalian Medical University,General Hospital of Northern Theater Command Training Base for Graduate,Shenyang 110016,China)

机构地区：[1]中国医科大学北部战区总医院研究生培养基地,辽宁沈阳110016 [2]北部战区总医院肝胆外科,辽宁沈阳110016 [3]北部战区总医院卫勤部,辽宁沈阳110016 [4]大连医科大学北部战区总医院研究生培养基地,辽宁沈阳110016

出　　处：《临床军医杂志》2025年第2期114-118,123,共6页Clinical Journal of Medical Officers

基　　金：辽宁省科技计划联合计划(2024JH2/102600291)。

摘　　要：目的开发一种基于瘤内和瘤周区域的计算机断层扫描放射组学特征联合临床因素的机器学习模型术前预测肝细胞癌(HCC)磷脂酰基醇蛋白聚糖3(GPC3)的表达。方法回顾性分析自2017年1月至2024年6月北部战区总医院肝胆外科收治的接受根治性手术的165例HCC患者的临床资料,按照7∶3比例将其随机分为训练集(n=114)与验证集(n=51)。使用Python软件分别提取瘤内和瘤周3 mm区域的放射组学特征,通过组内相关系数、t检验、皮尔逊相关系数、最小绝对收缩和选择算子回归法选择与HCC的GPC3表达相关的放射组学特征。采用逻辑回归算法构建瘤内、瘤周和融合模型,比较和选择最佳模型。为进一步提升模型的预测能力,使用最佳模型放射组学标签和经单因素和多因素Logistic回归筛选的临床风险因素构建联合模型,并绘制列线图。采用受试者工作特征曲线(ROC)和校准曲线来评估各模型的预测性能,用决策曲线来评估临床效益。结果165例HCC患者中,83例患者病理诊断为GPC3阳性设为GPC3阳性组,82例患者病理诊断为GPC3阴性设为GPC3阴性组。GPC3阳性组的甲胎蛋白(AFP)水平及乙肝病毒感染比例均高于GPC3阴性组,肝硬化比例低于GPC3阴性组,差异均有统计学意义(P<0.05)。训练集和验证集临床特征资料比较,差异均无统计学意义(P>0.05)。在训练集和验证集中,联合融合模型放射组学标签和AFP水平的临床-放射组学模型预测性能最高(曲线下面积分别为0.876、0.815)。校准曲线的霍斯默-莱梅肖拟合优度检验显示P>0.05(训练集:P=0.286、0.808、0.647;验证集:P=0.127、0.169、0.370),表明各模型的预测和实际GPC3表达状态之间有良好的一致性。决策曲线表明,使用列线图来评估GPC3表达状态的总体净效益大于临床模型和放射组学模型。结论联合瘤内和瘤周区域放射组学特征和AFP的机器学习模型能在术前准确预测HCC患者�Objective To develop a machine learning model for preoperative prediction of phosphatidyl acylolproteoglycan 3(GPC3)expression in hepatocellular carcinoma(HCC)based on intratumor and peritumoral radiomic features of computed tomography combined with clinical factors.Methods The clinical data of 165 HCC patients admitted to the Department of Hepatobiliary Surgery,General Hospital of Northern Theater Command from January 2017 to June 2024 were retrospectively analyzed,and they were randomly divided into a training set(n=114)and a validation set(n=51)according to a ratio of 7∶3.Python software was used to extract the radiomic features of the intra-and peri-tumor 3 millimeters region,respectively,and the radiomic features associated with GPC3 expression of HCC were selected by the intra-group correlation coefficient,t test,Pearson correlation coefficient,minimum absolute contraction and selection operator regression.The intratumoral,periatumoral and fusion models were constructed by logistic regression algorithm,and the best models were compared and selected.To further improve the predictive power of the model,a combined model was constructed using the best model radiomic label and clinical risk factors screened by univariate and multivariate Logistic regression,and a column graph was drawn.Receiver operating characteristic curve(ROC)and calibration curve were used to evaluate the predictive performance of each model,and decision curve was used to evaluate clinical benefit.Results Among the 165 HCC patients,83 were pathologically diagnosed as GPC3 positive and 82 were pathologically diagnosed as GPC3 negative.The level of alpha-fetoprotein(AFP)and the proportion of hepatitis B virus infection in GPC3-positive group were higher than those in GPC3-negative group,and the proportion of cirrhosis was lower than those in GPC3-negative group,with statistical significance(P<0.05).There was no significant difference in clinical characteristics between the training set and the verification set(P>0.05).In both the training se

关 键 词：肝细胞癌 磷脂酰基醇蛋白聚糖3 放射组学 机器学习 

分 类 号：R735.7[医药卫生—肿瘤]