基于术前MRI的影像组学机器学习模型预测软组织肿瘤Ki-67增殖指数  

Preoperative MRI-based radiomics machine learning model in predicting Ki-67 proliferation index in soft tissue neoplasms

作　　者：连欣怡 李杰[1,2,3] 付海花 刘松 刘吉华[1,2,3] 黄勇华 王鹤翔[1,2,3] LIAN Xinyi;LI Jie;FU Haihua;LIU Song;LIU Jihua;HUANG Yonghua;WANG Hexiang(Department of Radiology,the Affiliated Hospital of Qingdao Uhniversity,Qingdao,Shandong Province 266003,China;Health Bureau of Linqu County,Linqu,Shandong Province 262600,China;Department of Radiology,Puyang Oilfield General Hos pital,Puyang,He'nan Province 457001,China)

机构地区：[1]青岛大学附属医院放射科,山东青岛266003 [2]临胸县卫生健康局,山东临胸262600 [3]濮阳市油田总医院放射科,河南濮阳457001

出　　处：《实用放射学杂志》2022年第6期963-967,986,共6页Journal of Practical Radiology

摘　　要：目的建立并验证基于术前MRI影像组学机器学习模型预测软组织肿瘤Ki-67增殖指数的表达情况.方法将177例患者随机分为训练集(n=110)和验证集(n=67).利用最小冗余最大相关(mRMR)和LASSO算法作为特征选择方法,建立基于决策树(DT)、逻辑回归(LR)、随机森林(RF)、支持向量机(SVM)、自适应增强(AdaBoost)5种分类器的机器学习模型,构建影像组学标签.根据临床数据和MRI特征,建立临床模型.结合影像组学标签和临床模型,建立影像组学诺模图.通过受试者工作特征(ROC)曲线的曲线下面积(AUC)评估模型的预测性能.结果临床模型在训练集和验证集中AUC分别为0.689[95%置信区间(CI)0.593~0.784]和0.581(95%CI 0.454~0.708).5种机器学习模型中最佳的机器学习模型是mRMR-LASSO(特征选择)+SVM(分类器),在训练集和验证集中预测Ki-67高表达的AUC分别为0.998(95%CI 0.995~1.000)和0.798(95%CI 0.671~0.926).影像组学诺模图在训练集和验证集中,AUC分别为0.911(95%CI 0.854~0.967)和0.738(95%CI 0.609~0.867).临床模型和诺模图的AUC均低于机器学习模型.结论基于MRI的影像组学机器学习模型在预测软组织肿瘤Ki-67表达方面具有较好的性能,有潜力成为一种无创方法实现术前对Ki-67增殖指数的预测.Objective To develop and validate a preoperative MRI-based radiomics machine learning model so as to predict the expression of Ki-67 proliferation index in soft tissue neoplasms.Methods A total of 177 patients were randomly divided into a training set(n=110)and a validation set(n=67).The minimum redundancy maximum relevance(mRMR)and LASSO algorithm was used for feature selection.F urthermore,five machine learning models including decision tree(DT),logistic regression(LR),random forest(RF),support vector machine(SVM)and adaptive boosting(AdaBoost)classifier were established.The clinical model was established according to the clinical data and MRI features.Combined with the radiomics signature and clinical model,the radiomics nomogram was established.Areas under the curve(AUC)of the receiver operating characteristic(ROC)were used to evaluate the predictive performance of the model.Results The AUC of clinical model were 0.689[95% confidence interval(CI)0.593-0.784]and 0.581(95%CI 0.454-0.708)in the training set and the validation set,respectively.The best machine learning model was mRMR-LASSO(feature selection)+SVM(classifier),and the AUC for predicting the Ki-67 high expression in the training set and the validation set were 0.998(95%CI 0.995-1.000)and 0.798(95%CI 0.671-0.926),respectively.In the training set and the validation set,the AUC of radiomics nomogram was 0.911(95%CI 0.854-0.967)and 0.738(95%CI 0.609-0.867),respectively.The AUC of both the clinical model and the nomogram was lower than that of the machine learning model.Conclusion The MRI based radiomics machine learning model presents great performance in predicting the expression of Ki-67 in soft tissue neoplasms and has the potential to be a non invasive preoperative method for predicting Ki-67 proliferation index.

关 键 词：磁共振成像 软组织肿瘤 影像组学 人工智能 

分 类 号：R445.2[医药卫生—影像医学与核医学] R738.6[医药卫生—诊断学] TP18[医药卫生—临床医学]