基于MRI的生境影像组学预测子宫内膜癌分子亚型的双中心临床研究  

MRI-based habitat radiomics analysis for identifying molecular subtypes of endometrial cancer:a feasible study from two institutions

作　　者：金文韬 王添平 陈晓军 张国福 李海明 张鹤 JIN Wen-tao;WANG Tian-ping;CHEN Xiao-jun;ZHANG Guo-fu;LI Hai-ming;ZHANG He(Department of Radiology,Obstetrics and Gynecology Hospital,Fudan University,Shanghai 200011,China;Department of Gynecological Oncology,Obstetrics and Gynecology Hospital,Fudan University,Shanghai 200011,China;Department of Radiology,Fudan University Shanghai Cancer Center,Shanghai 200032,China)

机构地区：[1]复旦大学附属妇产科医院放射科,上海200011 [2]复旦大学附属妇产科医院妇科肿瘤科,上海200011 [3]复旦大学附属肿瘤医院放射科,上海200032

出　　处：《复旦学报（医学版）》2024年第6期890-899,共10页Fudan University Journal of Medical Sciences

摘　　要：目的建立基于MRI术前子宫内膜癌(endometrial cancer,EC)的分子亚型的生境影像组学预测模型。方法回顾性收集2家医学中心经病理证实的EC患者,分别纳入训练组(n=270)和测试组(n=70)。所有患者均进行了术前MRI及病理组织学和分子亚型诊断。首先根据扩散加权成像(diffusion-weighted imaging,DWI)和对比增强(contrast enhancement,CE)图像对肿瘤进行生境亚区域分区,随后从T1加权成像(T1-weighted imaging,T1WI)、T2加权成像(T2-weighted imaging,T2WI)、DWI和CE图像的不同亚区域提取生境影像组学特征。应用3种机器学习分类器,包括逻辑回归、支持向量机和随机森林,分别建立预测p53异常型EC的模型并进行效能验证,表现出最佳综合预测性能的模型被选为生境影像组学模型。采用相同程序,建立基于T1WI、T2WI、DWI和CE共4个序列的全区域影像组学模型及临床模型。采用受试者工作特性曲线评估模型的效能,使用DeLong检验比较模型的差异。使用决策曲线分析评价模型应用的临床收益。结果经特征选择后保留8个生境影像组学特征建立生境影像组学模型、10个全区域影像组学特征建立影像组学模型和3个临床特征建立临床模型。生境影像组学模型曲线下面积(area under the curve,AUC)最高,分别为0.855(0.788~0.922,训练集)和0.769(0.631~0.907,验证集)。DeLong检验显示训练集的生境影像组学模型效能优于全区域影像组学模型(P=0.001),但测试集差异不显著(P=0.543);两组生境影像组学模型效能均优于临床模型(P=0.007,训练集;P=0.038,验证集)。DCA曲线显示该模型在阈值概率0.2~0.8之间均可对临床诊断提供收益。结论基于MRI的生境影像组学模型可以较准确地预测p53异常型的EC,效能优于全区域影像组学和临床模型,有助于术前EC的无创性分子亚型分型。Objective To develop an MRI-based habitat radiomics model for the preoperative prediction of endometrial cancer(EC)molecular subtypes.Methods Patients with pathologically proven EC from two hospitals were included in the training(n=270)and testing(n=70)cohorts.All patients had preoperative MRI and histological and molecular diagnoses.First,the tumor was divided into habitat subregions based on diffusion-weighted imaging(DWI)and contrast-enhanced(CE)images.Subsequently,habitat radiomic features were extracted from different subregions of T1-weighted imaging(T1WI),T2-weighted imaging(T2WI),DWI,and CE images.Three machine learning classifiers,including logistic regression,support vector machines,and random forests,were applied to develop predictive models for p53-abnormal endometrial cancer,with model performance validated.The model demonstrating the best overall predictive performance was selected as the habitat radiomics model.Using the same procedure,a whole-region radiomics model based on T1WI,T2WI,DWI,and CE sequences and a clinical model were constructed.The performance of the models was evaluated using receiver operating characteristic curves,and DeLong’s test was employed to compare differences between the models.Decision curve analysis was used to assess the clinical benefits of the models’application.Results After feature selection,eight habitat radiomic features were retained to construct the habitat radiomics model,ten features for the whole-region radiomics model,and three clinical features for the clinical model.The habitat radiomics model achieved the highest area under the curve(AUC),with 0.855(0.788-0.922)in the training cohort and 0.769(0.631-0.907)in the testing cohort.DeLong’s test showed that the habitat radiomics model outperformed the whole-region radiomics model in the training cohort(P=0.001),but there was no significant difference in the testing cohort(P=0.543).In both cohorts,the habitat radiomics model outperformed the clinical model(P=0.007,training cohort;P=0.038,testing cohort).D

关 键 词：子宫内膜癌(EC) MRI 预测模型 生境 影像组学 

分 类 号：R730.44[医药卫生—肿瘤] R737.33[医药卫生—临床医学]