MRI影像组学鉴别肝原位癌和肝内胆管细胞癌的应用研究  

Application of MRI Imaging Omics to Distinguish Carcinoma in Situ of Liver and Intrahepatic Cholangiocarcinoma

作　　者：张苏雅 张帅[1] 吴佳妮 黄志成[1] ZHANG Suya;ZHANG Shuai;WU Jia´ni;HUANG Zhicheng(Department of Radiology,Jilin Provincial Cancer Hospital,Changchun,Jilin Province,130012 China)

机构地区：[1]吉林省肿瘤医院放射科,吉林长春130021

出　　处：《世界复合医学》2023年第2期14-18,共5页World Journal of Complex Medicine

基　　金：国家癌症中心攀登基金(NCC201907B04);吉林省科技发展计划项目(20210203092SF)。

摘　　要：目的探讨基于常规MRI腹部检查图像的影像组学与机器学习相结合在鉴别诊断肝原位癌和胆管细胞癌中的应用价值。方法回顾性分析2020年10月—2021年10月吉林省肿瘤医院经术后病理证实为肝原位癌和肝内胆管细胞癌患者134例的临床资料,所有患者术前均接受磁共振肝脏平扫检查,根据病理结果将肝原位癌患者作为第1组,肝内胆管细胞癌患者作为第2组,比较两组性别和年龄差异。采用图像纹理特征提取软件MaZda(Version 4.6)沿着磁共振影像中肿瘤病灶边缘勾画感兴趣区域(ROI),提取MRI病灶区域包括直方图、灰度游程矩阵、小波变换在内的纹理特征参数。采用主成分分析(principal component analysis,PCA)对数据进行降维,保留两组间差异明显的影像组学特征,用以构建机器学习和诊断模型。按照5折交叉检验的方式将数据集分为训练组和验证组,采用Logistic机器学习算法对数据集进行处理,以构建肝原位癌和肝内胆管细胞癌的鉴别诊断模型,并获得该模型在5倍交叉验证中的诊断效能参数,包括AUC、准确度、特异度和敏感度。结果第1组肝原位癌81例,第2组肝内胆管细胞癌53例。共提取得到病灶最大层面320个影像组学特征参数,经过数据降维和主成分分析,最终保留组间差异明显的6个影像组学参数以构建肝原位癌和肝内胆管细胞癌的预测模型。结果显示,逻辑回归模型(Logistic)在5折交叉验证中的具体参数AUC为(0.81±0.05),其准确率、特异度、敏感度分别为0.76、0.78、0.74。结论基于MRI影像组学建立的机器学习模型可以用于鉴别诊断肝原位癌和肝内胆管细胞癌,诊断准确率为0.76。Objective To explore the application value of combining imaging omics and machine learning based on conventional MRI abdominal images in the differential diagnosis of carcinoma in situ of liver and intrahepatic cholangiocarcinoma.Methods The clinical data of 134 patients with carcinoma in situ of liver and intrahepatic cholangiocarcinoma confirmed by postoperative pathology in Jilin Provincial Cancer Hospital from October 2020 to October 2021 were retrospectively analyzed.All patients received MRI liver plain scan before surgery.Carcinoma in situ of liver was divided into group 1 and intrahepatic cholangiocarcinoma into group 2 according to pathological results.Gender and age differences between the two groups were compared.Image texture feature extraction software MaZda(Version 4.6)was used to outline the ROI along the edge of tumor lesions in MRI images,and the texture feature parameters of MRI lesions including histogram,gray run matrix and wavelet transform were extracted.principal component analysis(PCA)was used to reduce the dimensionality of the data and retain the image omics features with significant differences between the two groups for the construction of machine learning and diagnosis models.The data sets were divided into training group and validation group according to the method of 5-fold cross test,and the Logistic machine learning algorithm was used to process the data sets,so as to construct the differential diagnosis model of carcinoma in situ of liver and intrahepatic cholangiocarcinoma.The diagnostic efficacy parameters,including AUC,accuracy,specificity and sensitivity,were obtained in 5-fold cross validation.Results There were 81 cases of carcinoma in situ of liver in group 1,53 cases of intrahepatic cholangiocarcinoma in group 2.A total of 320 imaging characteristics were extracted from the maximum focal layer.After data reduction and principal component analysis,6 imaging parameters with significant differences between groups were retained to construct the prediction models of carcinoma

关 键 词：肝原位癌 肝内胆管细胞癌 磁共振成像 影像组学 特征提取 机器学习 

分 类 号：R4[医药卫生—临床医学]