集成机器学习构建透明细胞肾细胞癌免疫影像分型及预测患者生存的价值  

Immune⁃radiomic phenotype based on ensemble machine learning in predicting survival of clear cell renal cell carcinoma

作　　者：李巧 王宇昊[2] 夏一凡 张玉东[1] LI Qiao;WANG Yuhao;XIA Yifan;ZHANG Yudong(Department of Radiology,the First Affiliated Hospital of Nanjing Medical University,Nanjing 210029,China;Department of Urology,the First Affiliated Hospital of Nanjing Medical University,Nanjing 210029,China)

机构地区：[1]南京医科大学第一附属医院放射科,江苏南京210029 [2]南京医科大学第一附属医院泌尿外科,江苏南京210029

出　　处：《南京医科大学学报（自然科学版）》2023年第9期1265-1272,共8页Journal of Nanjing Medical University(Natural Sciences)

基　　金：国家自然科学基金(82272082)。

摘　　要：目的:基于集成机器学习(machine learning,ML)及增强CT构建透明细胞肾细胞癌(clear cell renal cell carcinoma,ccRCC)免疫影像(immuno-radiomics,ImRad)分型并探讨其对患者预后生存的预测价值。方法:收集癌症基因图谱(the cancer genome atlas,TCGA)数据库113例ccRCC患者的临床、影像及基因表达信息,提取全肿瘤影像组学特征,基于集成ML进行特征筛选并构建免疫浸润、肿瘤突变负荷、免疫耗竭相关基因的ImRad分型。进一步多因素Cox回归分析ImRad对患者总生存(overall survival,OS)的预测效能。结果:经ML构建30个ImRad分型,经五折法验证,基于朴素贝叶斯算法对肿瘤免疫微环境的预测效能最佳(曲线下面积0.717~0.956)。与基于临床、病理及ImRad单模态指标对比,融合临床-病理及ImRad的模型预测OS的效能最佳。ImRad特征中,Rad-激活态肥大细胞等8个特征是OS的独立预测因子。结论:基于集成ML及CT组学分析可预测ccRCC免疫微环境并提高预测患者术后生存的效能。Objective:To construct hybrid immune⁃radiomic(ImRad)phenotypes of clear cell renal cell carcinoma(ccRCC)based on ensemble machine learning(ML)and contrast⁃enhanced CT,and to investigate its predictive value for survival.Methods:The clinical,CT imaging and gene⁃expression information of 113 ccRCC patients were collected from TCGA database.Radiomic features were extracted from whole tumor.ImRad predictors were constructed on tumor immune infiltration,tumor mutational burden,immune exhaustion gene expression after feature selection based on ensemble ML.Further,the predictive value of ImRad for overall survival(OS)was assessed using multivariate Cox regression analysis.Results:Among the 30 ImRad contructed by ensemble ML and validated by 5 folds cross validation,Naive Bayes algorithm achieved the generally best performance(area under the curve:0717-0.956).Clinicopathologic immune comprehensive model on predicting OS achieved the best performance,surpassing single modality indicators based on clinical,pathological and ImRad.Among ImRad features,Rad⁃Mast_cells_activated was one of eight independent predictive factors for patients’prognosis.Conclusion:CT radiomics based on ensemble ML can predict immune microenvironment and improve the prediction efficiency of postoperative survival of ccRCC.

关 键 词：透明细胞肾细胞癌 CT影像组学 机器学习 免疫浸润 

分 类 号：R737.11[医药卫生—肿瘤]