CT影像特征预测胸腺上皮性肿瘤病理分型  

Prediction of pathological classification of thymic epithelial tumors based on CT imaging features

作　　者：赵静[1] 黄晓媚 杨俊[3] 罗振东 曾伟雄 张妮[1] 秦耿耿[1] 文戈[1] ZHAO Jing;HUANG Xiaomei;YANG Jun;LUO Zhendong;ZENG Weixiong;ZHANG Ni;QIN Genggeng;WEN Ge(Department of Radiology,Nanfang Hospital,Southern Medical University,Guangzhou 510515,China;Department of Medical Imaging Education and Research,Nanfang Hospital,Southern Medical University,Guangzhou 510515,China;Department of Radiology,The Tenth Affiliated Hospital of Southern Medical University,Dongguan 523050,China;Department of Radiology,the University of Hong Kong-Shenzhen Hospital,Shenzhen 518000,China)

机构地区：[1]南方医科大学南方医院影像诊断科,广东广州510515 [2]南方医科大学南方医院医学影像教研室,广东广州510515 [3]南方医科大学第十附属医院放射科,广东东莞523050 [4]香港大学深圳医院放射科,广东深圳518000

出　　处：《分子影像学杂志》2024年第8期836-843,共8页Journal of Molecular Imaging

摘　　要：目的探讨基于临床及CT影像特征构建机器学习模型预测胸腺上皮性肿瘤病理分型并评估其预测效能。方法回顾性收集2006年1月~2023年6月在南方医科大学南方医院经病理证实的221例胸腺上皮性肿瘤患者资料,包括临床信息、CT影像和病理结果,并根据简化病理分型将患者分为低危型(A、AB、B1型)和高危型(B2、B3、胸腺癌)。以7:3的比例将患者随机划分为训练集(n=159)和验证集(n=62);在训练集中,采用单因素Logistic回归分析低危组和高危组临床及CT特征的差异性,应用逐步回归和LASSO回归进一步降维筛选特征,构建4种机器学习模型(Logistic回归、随机森林、决策树和支持向量机模型)。在验证集中通过ROC曲线分析曲线下面积评价模型预测效能。结果221例胸腺上皮性肿瘤患者中105例为低危型(训练集74例,验证集31例),116例为高危型(训练集85例,验证集31例)。单因素分析结果显示,高危组与低危组比较性别、胸痛差异有统计学意义(P<0.05);通过逐步回归法选取3个CT影像特征(肿瘤强化程度、心包或大血管侵犯、胸膜侵犯)构建多因素Logistic回归模型,通过LASSO回归分析最终筛选出8个临床及CT影像特征构建随机森林、决策树及支持向量机模型。模型在训练集上的曲线下面积分别为0.793、0.854、0.761、0.816,在验证集上的曲线下面积分别为0.819、0.742、0.710、0.811。结果表明Logistic回归模型泛化性优于其他3个模型。结论基于CT影像特征构建的Logistic回归模型在预测胸腺上皮性肿瘤病理分型上具有良好的诊断效能,有望协助临床早期无创性识别高危型胸腺瘤及胸腺癌。Objective To investigate the construction and predictive performance of machine learning models based on clinical and CT imaging features for predicting pathological subtypes of thymic epithelial tumors(TETs).Methods This retrospective study included data from 221 patients with pathologically confirmed TETs at Nanfang Hospital,Southern Medical University,between January 2006 and June 2023.The data collected included clinical information,CT images,and pathological results.According to simplified pathological classification,the patients were classified into low-risk group(type A,AB,B1)and highrisk group(type B2,B3,thymic carcinoma).The included cases were randomly divided into the training set(n=159)and the validation set(n=62)at a ratio of 7:3.In the training set,univariate logistic regression was used to analyze the differences of clinical and CT characteristics between low-risk group and high-risk group.Feature selection was performed using stepwise regression and LASSO regression to construct four machine learning models,that were logistic regression,random forest,decision tree,and support vector machine.Model performance was evaluated in the validation set by AUC.Results Among 221 cases of thymic epithelial tumors,105 cases were low-risk type(74 in training set,31 in validation set)and 116 cases were high-risk type(85 in training set,31 in validation set).The results of univariate analysis showed that there were significant differences in sex and chest pain between high-risk group and low-risk group(P<0.05).Three CT features(tumor enhancement,pericardial or great vessel invasion,and pleural invasion)were selected using stepwise regression to construct a multivariate logistic regression model.Eight clinical and CT imaging features were selected through LASSO regression analysis for constructing random forest,decision tree,and support vector machine models.The AUCs for the models in the training set were 0.793,0.854,0.761,and 0.816,and in the validation set,they were 0.819,0.742,0.710,and 0.811,respectively.Thes

关 键 词：胸腺上皮性肿瘤 影像学特征 临床特征 预测模型 

分 类 号：R736.3[医药卫生—肿瘤] R730.44[医药卫生—临床医学]