基于MRI肿瘤和瘤周的机器学习模型预测直肠癌新辅助治疗后淋巴结转移的研究  

Prediction of N⁃staging of Rectal Cancer After Neoadjuvant Therapy Based on T_(2)WI Tumor and Mesorectal Machine Learning

作　　者：杨倩 郭小芳[1] 袁子龙[1] 聂婷婷 刘玉林[1] YANG Qian;GUO Xiaofang;YUAN Zilong(Department of Radiology,Hubei Cancer Hospital,Wuhan,Hubei Province 430079,P.R.China)

机构地区：[1]湖北省肿瘤医院放射科,湖北省结直肠癌临床医学研究中心,武汉市结直肠癌临床医学研究中心,430079

出　　处：《临床放射学杂志》2024年第8期1371-1378,共8页Journal of Clinical Radiology

基　　金：国家癌症中心攀登基金临床研究重点课题项目(编号:NCC201917B05);湖北省肿瘤医院生物医学中心专项科研基金项目(编号:2022SWZX06)。

摘　　要：目的探讨基于T_(2)WI肿瘤和直肠系膜的机器学习对直肠癌新辅助治疗后淋巴结转移的预测价值。方法回顾性分析171例直肠癌病例新辅助治疗前的临床影像资料及术后病理资料。依据术后病理结果分为N-(N0)和N+(N1、N2)。统计学分析筛选出与N分期(pN)存在相关性的临床特征(实验室指标及MRI影像学表现)。使用ITK⁃SNAP软件在新辅助治疗前直肠MRI高分辨率T_(2)WI图像上,手动绘制直肠癌整个瘤灶作为感兴趣区(ROI⁃1),直肠系膜区域作为ROI⁃2;从ROI⁃1、ROI⁃2提取所有影像组学特征,保留稳定性较好(ICC≥0.75)的特征。采用最小绝对紧缩与选择算子(LASSO)方法从ROI⁃1、ROI⁃2及融合特征(ROI⁃1+ROI⁃2+临床特征)中筛选出与pN最相关的特征。将筛选得到的四组特征,分别采用支持向量机(SVM)、K最邻近算法(KNN)、随机森林(RF)、极端随机树(ET)、梯度提升决策树(XGBoost)、LightGBM(LGBM)、逻辑回归(LR)七种机器学习算法构建pN的预测模型,使用受试者工作特征曲线(ROC)评估模型的性能。结果171例直肠癌患者,手术结果显示N-组92例,N+组79例。ROI⁃1筛选得到的特征构建的7种模型中,LR模型效能最佳,测试集中受试者工作特征曲线曲线下面积(AUC)0.656、准确性0.714、敏感度0.583、特异度0.783。ROI⁃2中,SVM模型效能最佳,测试集中,AUC、准确性、敏感度、特异度相应为0.721、0.657、0.917、0.522。临床模型中,LR模型效能最佳,相应为0.768、0.771、0.833、0.773。融合模型中,LR模型预测效能最佳,相应为0.866、0.800、0.917、0.739。结论基于机器学习的直肠肿瘤及直肠系膜MRI影像组学分析均能预测直肠癌新辅助治疗后淋巴结转移情况,融合双区域多组学特征采用逻辑回归方法可提高预测模型的准确性。Objective To explore the predictive value of machine learning based on T2WI tumor and mesorectum for lymph node metastasis of rectal cancer after neoadjuvant therapy.Methods A retrospective analysis was conducted on the clinical imaging data and postoperative pathological data of 171 cases of rectal cancer before neoadjuvant therapy.According to postoperative pathological results,it was divided into N⁃(N0)and N+(N1,N2).Statistical analysis screened clinical features(laboratory indicators and MRI imaging findings)that were correlated with N staging(pN).The ITK⁃SNAP soft-ware was used to manually draw the whole tumor focus of rectal cancer as the region of interest(ROI⁃1)and the mesorectal region as the ROI⁃2 on the high⁃resolution T2WI image of rectal MRI before the new adjuvant treatment;Extract all imaging omics features from ROI⁃1 and ROI⁃2,retaining features with good stability(ICC≥0.75).Using the Least Absolute Shrinkage and Selection Operator(LASSO)method,the most relevant features to pN were selected from ROI⁃1,ROI⁃2,and fusion features(ROI⁃1+ROI⁃2+clinical features).The four groups of features selected were used to construct pN prediction models using seven machine learning algorithms,namely,support vector machine(SVM),K⁃nearest neighbor algorithm(KNN),Random forest(RF),extreme Random tree(ET),gradient lifting decision tree(XGBoost),LightGBM(LGBM),and Logistic regression(LR).Results Among the 171 rectal cancer patients,the surgical results showed 92 cases in the N⁃group and 79 cases in the N+group.Among the seven models constructed based on ROI⁃1 screening,the LR model had the best performance,with AUC of 0.656,accuracy of 0.714,sensitivity of 0.583,and specificity of 0.783 in the test set.In ROI 2,the SVM model had the best performance,with corresponding AUC,accuracy,sensitivity,and specificity of 0.721,0.657,0.917,and 0.522 in the test set.In clinical models,the LR model had the best performance,with corresponding values of 0.768,0.771,0.833,and 0.773.In the fusion model,the

关 键 词：磁共振成像 影像组学 直肠癌 机器学习 N分期 

分 类 号：R735.37[医药卫生—肿瘤]