不同机器学习方法构建MRI影像组学联合临床指标预测肝细胞癌患者射频消融术后早期复发模型与评估
Construction and evaluation of different machine learning models based on MRI combined with clini-cal indicators for predicting early recurrence of patients with hepatocellular carcinoma after radiofre-quency ablation
李文华 1唐静 1王楠钧 2李雪萍 3王肖 1李天然1
作者信息
- 1. 中国人民解放军总医院第四医学中心放射诊断科,北京 100048
- 2. 中国人民解放军总医院第一医学中心消化内科医学部,北京 100853
- 3. 中国人民解放军总医院第一医学中心放射诊断科,北京 100853
- 折叠
摘要
目的 利用不同机器学习方法构建多模态MRI影像组学联合临床指标预测肝细胞癌患者射频消融术后早期复发的模型,并评估模型预测能力.方法 回顾性分析2015年1月至2021年12月中国人民解放军总医院第四医学中心和中国人民解放军总医院第一医学中心行射频消融治疗的肝细胞癌患者资料.共入组169例肝细胞癌患者,其中男性152例,女性17例,年龄(57.2±9.2)岁.按照8:2随机分为训练集(n=135)与测试集(n=34).训练集中复发49例,测试集中复发12例.基于训练集单因素和多因素logistic回归分析肝细胞癌患者射频消融术后早期复发的临床影响因素,应用方差阈值法、select K-best和LASSO回归依次筛选影像组学特征.采用支持向量机、logistic回归、随机森林三种机器学习分类器分别构建单纯影像组学或联合临床特征的术后早期复发预测模型,受试者工作特征(ROC)曲线评估模型的预测能力.结果 多因素logistic回归分析显示,术前甲胎蛋白＞20 μg/L、血小板计数＞140 ×109以及肿瘤位置特殊是肝细胞癌患者射频消融术后早期复发的影响因素(均P＜0.05).经方差阈值分析、select K-best、LASSO回归筛选出16个最优影像组学特征.采用支持向量机、logistic回归、随机森林分类器构建单纯影像组学预测肝细胞癌患者射频消融术后早期复发模型,测试集中预测肝细胞癌射频消融术后早期复发的ROC曲线下面积分别为0.826、0.830、0.826;构建的影像组学联合临床特征模型,测试集的ROC曲线下面积分别为0.830、0.830、0.909.测试集中,随机森林的ROC曲线下面积大于支持向量机和logistic回归(Z=2.19、3.98,P=0.008、0.008).结论 支持向量机、logisitic回归、随机森林三种学习方法基于临床指标及影像组学特征构建的联合模型预测肝细胞癌患者射频消融术后早期复发效能良好,其中,随机森林构建的模型最优.
Abstract
Objective To construct a model for predicting early recurrence of hepatocellular carci-noma(HCC)patients after radiofrequency ablation by different machine learning models based on multimo-dal MRI and clinical indicators,and to evaluate the predictive efficacy of the model.Methods The data of patients with HCC who underwent radiofrequency ablation in Fourth Medical Center of Chinese PLA General Hospital and the First Medical Center of Chinese PLA General Hospital from January 2015 to December 2021 were retrospectively analyzed.A total of 169 patients with HCC were enrolled,including 152 males and 17 females,aged(57.2±9.2)years.The training set(n=135)and the test set(n=34)were randomly di-vided according to 8:2.There were 49 cases recurrence in training set and 12 cases recurrence in test set.Based on the training set,the clinical influencing factors of early recurrence in patients with HCC after radio-frequency ablation were screened by univariated and multivariate logistic analysis,and the imaging features were sequentially screened by variance threshold method,select K-best and LASSO regression.Support vec-tor machine(SVM),logistic regression and random forest(RFOREST)were used to construct the predic-tion models of early postoperative recurrence with simple imagomics alone or combined clinical features,respectively,and the receiver operating characteristic(ROC)curve was used to evaluate the prediction efficiency of the models.Results Multivariate logistic regression analysis showed that preoperative alpha-fetoprotein＞20 μg/L,platelet count＞140 × 109 and tumor location were the influential factors for early recurrence of HCC patients after radiofrequency ablation(all P＜0.05).Through variance threshold analysis,select K-best and LASSO regression,16 optimal image omics features were selected.SVM,logistic regression and RFOREST were used to construct a simple imaging omics model for predicting early recur-rence of HCC patients after radiofrequency ablation.The areas under ROC curve of the test set were 0.826,0.830 and 0.826,respectively.And the areas under ROC curve of the constructed imagomics combined clinical model of test set were 0.830,0.830 and 0.909,respectively.The area under ROC curve of RFOREST in the test set was better than that of SVM and logistic regression(Z=2.19,3.98,P=0.008,0.008).Conclusion The combined model constructed by SVM,logistic regression and RFOREST based on clinical indicators and image omics features is effective in predicting the early recurrence of patients with HCC after radiofrequency ablation,and the model constructed by RFOREST is the best.
关键词
癌,肝细胞/射频消融术/复发/磁共振成像/影像组学Key words
Carcinoma,hepatocellular/Radiofrequency ablatio/Recurrence/Magnetic reso-nance imaging/Radiomics引用本文复制引用
基金项目
解放军总医院青年自主创新科学基金(22QNFC144)
出版年
2024
NETL
NSTL
万方数据