摘要目的:探讨基于增强CT图像构建影像组学模型及其与相关临床病理特征构建的融合模型在浸润性肺腺癌术前预测组织学分级中的应用价值。方法:回顾性队列研究。纳入2017年1月—2019年12月山西省肿瘤医院经手术病理确诊且组织学分级明确的313例浸润性肺腺癌患者，其中男175例、女138例，年龄35~81（60.1±8.3）岁。组织学分级：高级别（G3级）85例，低级别（G1、G2级）228例。313例患者术前均行胸部增强CT检查。患者按照7∶3的比例随机分为训练组219例（高级别59例、低级别160例）和验证组94例（高级别26例、低级别68例）。基于每位患者的术前增强CT图像各提取1 316个影像组学特征；训练组采用最大相关最小冗余、最小绝对收缩与选择算子算法和多因素logistic回归分析进行特征筛选及影像组学模型的构建。用单因素和多因素logistic回归分析从临床病理特征中筛选高级别浸润性肺腺癌的独立危险因素，联合影像组学模型建立融合模型，并绘制列线图。使用受试者操作特征曲线（ROC曲线）、校准曲线和决策曲线评估影像组学模型和融合模型对术前浸润性肺癌组织分级的预测效能及临床效益。结果:训练组与验证组患者的临床病理特征比较，差异均无统计学意义（ P值均>0.05）。训练组和验证组内的高级别与低级别患者比较，吸烟史、肿瘤实性程度、肿瘤最大径、血清癌胚抗原和细胞角蛋白19片段水平的差异均有统计学意义（ P值均<0.05），年龄、毛刺征、分叶、胸膜牵拉的差异均无统计学意义（ P值均>0.05）；此外，验证组的低级别与高级别患者性别分布比较，差异有统计学意义（ χ2=4.70， P=0.030）。特征降维后得到5个影像组学特征用于构建影像组学模型，影像组学模型在训练组和验证组ROC曲线的曲线下面积（AUC）分别为0.852（95% CI：0.734~0.882、0.837（95% CI：0.733~0.863）。单因素和多因素logistic回归分析筛选出高级别浸润性肺腺癌的独立危险因素有吸烟史、肿瘤实性程度、肿瘤最大径、血清癌胚抗原和细胞角蛋白19片段水平，其与影像组学模型共同构建的融合模型在训练组和验证组ROC曲线的AUC分别为0.888（95% CI：0.721~0.894）、0.876（95% CI：0.707~0.852）。校准曲线分析显示，影像组学模型和融合模型均有良好的校准性能。决策曲线分析显示，2种预测模型均有一定的临床效益，其中融合模型净收益值更大。 结论:基于术前增强CT图像构建的影像组学模型及其与高级别浸润性肺腺癌的独立危险因素结合建立的融合模型，对于浸润性肺腺癌的组织学分级均有良好的术前预测价值，并且后者的预测效能相对更高。

abstractsObjective:This study aimed to explore the value of preoperative prediction radiomics fusion model based on contrast-enhanced CT images and related clinicopathogical features for predicting pathological grades in patients with invasive lung adenocarcinoma.Methods:A retrospective cohort study was conducted. The clinicopathological information of 313 patients with invasive lung adenocarcinoma, who underwent preoperative contrast-enhanced CT in Shanxi Provincial Cancer Hospital from January 2017, to December 2019, were obtained. All patients were diagnosed by surgical pathology, and they had clear pathological grades. The patients aged 35-81 (60.1±8.3) years, and they were composed of 175 males and 138 females. On the basis of pathological grading, the high-grade group (G3 adenocarcinoma) included 85 cases, and the low-grade group (G1 and G1 adenocarcinoma) included 228 cases. The 313 patients were randomly divided into training group (219, including 59 high-grade and 60 low-grade cases) and validation group (94, including 26 high-grade and 68 low-grade cases) at the ratio of 7∶3. A total of 1 316 radiomic features were extracted from the preoperative contrast-enhanced CT images. Then, a radiomics model was constructed using the maximum relevance minimum redundancy, the least absolute shrinkage, and the selection operator regression and multivariate logistic regression in the training group. The clinicopathological independent risk factors for high-grade invasive lung adenocarcinoma were selected using univariate and multivariate logistic regression. Then, a combined model integrating the radiomics model and the clinicopathological independent risk factors were established, and the corresponding nomogram were constructed. The predictive performance and clinical benefits of the radiomic and combined models were evaluated by analyzing the receiver operating characteristic curves, calibration curves, and decision curves.Results:No statistically significant differences were found when comparing the clinicopathological data of patients in the training and validation groups (all P values >0.05). Comparison of high-grade with low-grade patients within the training and validation groups showed that the differences in smoking history, tumor solidity, tumor maximum diameter, carcinoembryonic antigen, and cytokeratin 19 fragment were statistically significant (all P values <0.05), whereas the differences in age, spiculation, lobulation, and pleural retraction were not statistically significant (all P values >0.05). For gender comparison, the difference between low-grade and high-grade patients in the validation group was statistically significant ( χ2=4.70, P=0.030). Five key radiomics features were finally obtained after feature selection and used to establish the radiomics model. The area under the curve (AUC) of the radiomics model in the training and validation groups were 0.852 (95% CI:0.734-0.882) and 0.837 (95% CI:0.733-0.863), respectively. The independent clinical risk factors for high-grade invasive lung adenocarcinoma, as screened by univariate and multivariate logistic regression analyses, were smoking history, tumor solidity, tumor maximum diameter, carcinoembryonic antigen, and cytokeratin 19 fragment. The AUCs of the combined model that incorporated the radiomics model and above clinical risk factors were 0.888 (95% CI:0.721-0.894) and 0.876 (95% CI:0.707-0.852) in the training and validation groups, respectively. The calibration curves showed that the radiomics model and the combined model had good calibration. The decision curve analysis showed that the two predictive models had certain clinical importance, among which the combined model had the largest net profit. Conclusion:The radiomics model derived from preoperative enhanced CT images and the combined model integrating the radiomics model with the independent risk factors for high-grade invasive lung adenocarcinoma have good value in predicting the pathological grades of invasive lung adenocarcinoma before surgery, with the latter having relatively higher efficacy.