基于平扫MRI的影像组学及深度学习模型预测肝细胞癌微血管侵犯:一项双中心研究  

Radiomics and deep learning models based on unenhanced MRI to predict microvascular invasion in hepatocellular carcinoma:a two-center study

作　　者：张歌 钟淑媛 胡根文 李欣明[1] 全显跃[1] ZHANG Ge;ZHONG Shuyuan;HU Genwen;LI Xinming;QUAN Xianyue(Department of Radiology,Zhujiang Hospital of Southern Medical University,Guangzhou 510280,China;Department of Radiology,the Second Clinical Medical College of Ji’nan University,Shenzhen People’s Hospital,Shenzhen,Guangdong Province 518000,China)

机构地区：[1]南方医科大学珠江医院放射科,广东广州510280 [2]暨南大学第二临床医学院深圳市人民医院放射科,广东深圳518000

出　　处：《实用放射学杂志》2025年第3期424-428,共5页Journal of Practical Radiology

基　　金：广东省自然科学基金项目(2024A1515012170)。

摘　　要：目的 探讨基于平扫MRI的影像组学模型及深度学习模型在术前预测肝细胞癌(HCC)微血管侵犯(MVI)的价值.方法 回顾性选取2个中心经术后病理证实为HCC的189例患者,其中南方医科大学珠江医院119例作为训练集[60例MVI阴性,59例MVI阳性],深圳市人民医院70例作为外部测试集[38例 MVI阴性,32例 MVI阳性].通过单因素及多因素logistic回归分析临床指标,筛选出 MVI阳性的独立预测因素.采用深度迁移学习(DTL)方法和传统影像组学方法基于平扫的 MRI构建影像组学模型及深度学习模型,利用受试者工作特征(ROC)曲线以及曲线下面积(AUC)比较各模型预测效能,筛选最佳模型.DeLong检验比较各模型效能的差异.结果 碱性磷酸酶(ALP)及凝血酶原时间(PT)为 MVI阳性的独立预测因素(P<0.05).基于T2WI的深度学习模型预测效能最好,在训练集和外部测试集中的AUC分别为0.779[95%置信区间(CI)0.696~0.863]、0.741(95%CI 0.620~0.861),并且与基于T1WI的影像组学模型及临床模型之间有统计学差异(P<0.05).结论 基于T2WI的深度学习模型术前无创性预测HCC患者 MVI状态有一定的应用价值.Objective To explore the value of radiomics model and deep learning model based on unenhanced MRI in predicting microvascular invasion(MVI)of hepatocellular carcinoma(HCC)preoperatively.Methods A total of 189 patients with postoperative pathologically confirmed HCC from two centers were retrospectively selected,of which 119 cases from Zhujiang Hospital of Southern Medical University were used as the training set[60 cases with negative MVI,59 cases with positive MVI],and 70 cases from Shenzhen People’s Hospital were used as the external test set[38 cases with negative MVI and 32 cases with positive MVI].Clinical indicators were analyzed by univariate and multivariate logistic regression analysis and the independent predictors of positive MVI were screened.Deep transfer learning(DTL)and traditional radiomics methods were used to construct radiomics model and deep learning model based on unenhanced MRI.The predictive performances of each model were compared using receiver operating characteristic(ROC)curves and area under the curve(AUC).DeLong test was employed to compare statistical differences in performance of the models.Results Alkaline phosphatase(ALP)and prothrombin time(PT)were independent predictors of positive MVI(P<0.05).The deep learning model based on T2WI had the best predictive efficacy,with AUC of 0.779[95%confidence interval(CI)0.696-0.863]and 0.741(95%CI 0.620-0.861)in the training set and external test set,respectively,and there were statistically significant differences compared with the radiomics model and the clinical model based on T1WI(P<0.05).Conclusion Deep learning model based on T2WI has a certain application value in preoperative noninvasive prediction of MVI status in HCC patients.

关 键 词：肝细胞癌 微血管侵犯 影像组学 深度迁移学习 磁共振成像 

分 类 号：R735.7[医药卫生—肿瘤] R445[医药卫生—临床医学] R445.2