目的 构建一种基于CT影像组学的早期浸润性肺腺癌（ILAC）高级别病理成分的预测模型。 方法  收集2015年1月~2019年12月舟山医院胸心外科行肺癌根治术且术后病理诊断为Ⅰ期的共495例患者信息，包括性别、年龄、病理结果、肿瘤标志物和术前胸部CT影像等资料。将术后病理中的微乳头和实性成分定义为“高级别病理成分”，而不含高级别病理成分者归为低级别组，含高级别病理成分者归为高级别组。使用简单随机分组方法将纳入病例按7∶3比例随机分为训练集343例与验证集152例。通过科研平台对CT图像上结节的感兴趣区域进行逐层勾画并提取1950个影像组学特征。经F检验、Pearson相关系数和基于L1的特征选择等方法进行特征筛选，运用逻辑回归机器学习分类器构建模型，命名为模型2（mod 2），并计算获得影像组学评分（Radscore）。将一般信息和CT特征进行差异分析，将具有统计学意义的指标采用二元Logistic回归分析构建模型，命名为模型1（mod 1）。同时加入Radscore构建模型，命名为联合模型（comb mod）。计算三种模型的曲线下面积（AUC)、敏感性和特异性，绘制列线图。结果 共收集495例，其中训练集343例，验证集152例。性别、癌胚抗原、结节性质、病灶最大径等指标参与构建mod 1。影像组学特征中筛选获得12个特征用于建立mod 2。而comb mod表现最佳（训练集AUC：0.887，验证集AUC：0.875），且具有较好的临床实用性。结论 将一般CT特征和影像组学特征相结合所构建的模型，能较准确地预测早期ILAC中的高级别病理成分，为临床医生对患者手术方式的选择提供参考依据。 

 Objective To construct a predictive model for high-grade pathological components of early invasive lung adenocarcinoma(ILAC) based on radiomics.    Methods Collecting information on total 495 patients who underwent radical operation and were pathologically diagnosed as stage Ⅰ in the cardiothoracic surgery of Zhoushan Hospital from January 2015 to December 2019, including gender, age, pathological findings, tumor markers and preoperative chest CT images. The micropapillary and solid components in postoperative pathology were defined as “high-grade pathological components”, while those without high-grade pathological components were classified into the low-grade group and those with high-grade pathological components were classified into the high-grade group. And patients were randomly divided into the training set(343 cases) and the validation set(152 cases) with a ratio of 7∶3 using the simple randomization grouping method. The region of interest of nodules on CT images were delineated layer by layer by scientific research platform and 1950 radiomics features were extracted. And then those features were filtrated by Ftest, Pearson correlation coefficient, and L1 based feature selection. A model was built by using Logistic regression machine learning classifier, named mod 2, and radscore was also obtained. Differences between general information and CT features were analyzed. Binary Logistic regression analysis was used to construct a model for statistically significant variables, named mod 1. At the same time, Radscore was added to build the mod and named comb mod. The area under the curve(AUC), sensitivity and specificity of the three models were calculated. A nomogram was also drawn.   Results A total of 495 patients were divided into the training set (n=343) and the validation set (n=152). Gender, carcinoma embryonic antigen(CEA), nodule, and maximum diameter were screened out in clinical features and involved in constructing the mod 1. Twelve features were selected from the radiomics features to build mod 2. Comb mod performed best, training set AUC:0.887, validation set AUC:0.875, and had good clinical practicability.    Conclusion The model composed of general feature, CT feature and radiomics features could accurately predict high-grade pathological components in early ILAC, and provide references for clinicians to choose surgical method  for patients.