基于CT影像组学术前预测淋巴结阴性胃癌淋巴血管侵犯  

Preoperative Prediction of Lymphovascular Invasion of Node-Negative Gastric Cancer Based on CT Radiomics

作　　者：娄飞飞 陈青青 黄昊 王芳[1] 何杰 辛恩慧 胡红杰[1] LOU Feifei;CHEN Qingqing;HUANG Hao;WANG Fang;HE Jie;XIN Enhui;HU Hongjie(Department of Radiology,Sir Run Run Shaw Hospital,Zhejiang University School of Medicine,Hangzhou 310000,China;不详)

机构地区：[1]浙江大学医学院附属邵逸夫医院放射科,浙江杭州310000 [2]湖州市南浔区人民医院放射科,浙江湖州313000 [3]上海联影智能医疗科技有限公司,上海200030

出　　处：《中国医学影像学杂志》2024年第1期73-80,共8页Chinese Journal of Medical Imaging

基　　金：国家自然科学基金项目(82071988);浙江省重点研发计划(2019C03064);省部共建项目(WKJ-ZJ-1926)。

摘　　要：目的探讨基于CT影像组学术前预测淋巴结阴性胃癌淋巴血管侵犯(LVI)的价值,并结合临床变量构建列线图。资料与方法回顾性分析浙江大学医学院附属邵逸夫医院2019年1月—2021年6月173例淋巴结阴性且病理证实为胃癌患者的临床及CT影像,其中LVI阳性60例,LVI阴性113例,按7∶3随机分为训练组(n=121)和验证组(n=52)。基于训练组分别构建临床模型、影像组学模型、融合模型,并在验证组进行验证。临床资料和常规CT特征包括年龄、性别、肿瘤指标、肿瘤部位、肿瘤形态、强化幅度等,通过单因素及多因素分析筛选出临床显著变量并建立临床模型。用3D-Slicer软件勾画肿瘤感兴趣区并提取影像组学特征,用最小绝对值收缩和选择算子降维筛选特征,然后通过随机森林构建影像组学模型,并转化为随机森林评分。联合临床显著变量和随机森林评分构建融合模型并可视化为列线图。根据受试者工作特征曲线及曲线下面积(AUC)评估模型的预测效能,采用决策曲线分析评估临床实用性。结果影像组学模型优于临床模型,训练组和验证组中影像组学模型AUC分别为0.872(0.810~0.935)、0.827(0.707~0.947),临床模型AUC分别为0.767(0.682~0.852)、0.761(0.610~0.913)。列线图的预测效能得到进一步提高,AUC分别为0.898(0.842~0.953)、0.844(0.717~0.971)。决策曲线分析显示列线图的临床价值。结论本研究构建的影像组学模型可用于术前预测淋巴结阴性胃癌患者LVI状态,列线图可进一步提高预测效能。Purpose To explore the value of CT-based radiomics in the preoperative prediction of lymphatic invasion of node-negative gastric cancer,and to construct a nomogram combined with clinical variables.Materials and Methods The clinical and CT imaging data of 173 gastric cancer patients with lymph node negative and pathologically confirmed gastric cancer in the Sir Run Run Shaw Hospital from January 2019 to June 2021 were retrospectively analyzed.A total of 60 cases with lymphovascular invasion(LVI)positive patients and 113 cases with LVI negative patients were included,and randomly divided into train cohort(n=121)and test cohort(n=52)at 7∶3.Based on the train cohort,the clinical model,the radiomics model,the fusion model were constructed and verified in the test cohort.Clinical data and conventional CT features included age,gender,tumor marker,tumor location,tumor morphology,enhancement range,etc.The clinical significant variables were selected through univariate and multivariate analysis to establish the clinical model.The tumor regions of interest were segmented and radiomics features were extracted by using the 3D-Slicer software.Key features were screened through least absolute shrinkage and selection operator regression analysis,and then the radiomics model was constructed with random forest algorithm,and converted to random forest score(RF score).The fusion model was constructed via combining clinical significant variables and RF score,and visualized as a nomogram.The receiver operator characteristic curve and area under curve(AUC)were used to evaluate the prediction performance of the models.Decision curve analysis was used to calculate the clinical practicability.Results The radiomics model was superior to the clinical model.The radiomics model AUC of the train cohort and the test cohort were 0.872(0.810 to 0.935)and 0.827(0.707 to 0.947),the clinical model AUC were 0.767(0.682 to 0.852)and 0.761(0.610 to 0.913).The nomogram further improved the predictive efficiency,the AUC in train cohort and test cohort

关 键 词：胃肿瘤 影像组学 淋巴结阴性 淋巴血管侵犯 体层摄影术 X线计算机 

分 类 号：R735.2[医药卫生—肿瘤] R445.3[医药卫生—临床医学] R73-37