摘要
目的探讨基于机器学习的影像组学模型对子宫内膜癌患者无病生存期(disease-free survival,DFS)的预测价值。材料与方法回顾性分析双中心212例接受过根治性手术并正规随访的子宫内膜癌患者资料。提取所有患者T2WI序列中原发灶及瘤周5 mm区域的影像组学特征。采用5种机器学习方法(梯度提升机、最小绝对收缩和选择算法、随机生存森林、支持向量机和极端梯度上升)构建影像组学模型并计算最佳影像组学评分(radiomics score,Radscore)。分析Radscore对现有临床预测指标的增量价值并构建联合模型。最后,采用生物信息学分析揭示影像组学模型的生物学机制。结果基于梯度提升机的影像组学模型预测效能最佳,其在训练集和验证集预测1年、3年、5年DFS的曲线下面积(area under the curve,AUC)分别为0.977、0.986、0.995和0.745、0.764、0.802。多因素Cox回归分析显示临床分期、糖类抗原125(carbohydrate antigen 125,CA125)和Radscore为子宫内膜癌患者DFS的独立预测因素。联合模型在训练集和验证集预测1年、3年、5年DFS的AUC分别为0.926、0.894、0.864和0.828、0.839、0.873。同时,生信分析提示Radscore与子宫内膜癌患者免疫水平显著相关。结论基于机器学习的影像组学模型有助于子宫内膜癌DFS及免疫水平的预测。影像组学和临床指标的联合能进一步提高预测精度,为子宫内膜癌患者的预后预测和个体化治疗提供参考依据。
Objective:To investigate the predictive value of machine learning-based radiomics model for disease-free survival(DFS)in endometrial cancer patients.Matirials andMethods:Data from 212 endometrial cancer patients who had undergone radical surgery in a dual-center were retrospectively analyzed.Radiomics features of tumor and peri-tumor 5 mm region in T2WI sequences were extracted for all patients.Five machine learning methods(gradient boosting machines,the least absolute shrinkage and selection operator,random survival forest,support vector machine,and extreme gradient boosting)were used to construct the radiomics model and calculate the best radiomics score(Radscore).The incremental value of Radscore to existing clinical predictors was analysed and a combined model was constructed.Finally,bioinformatics analysis was used to reveal the biological mechanisms of the radiomics models.Results:The combined radiomics model based on gradient boosting machines showed the best predictive efficacy,with AUC of 0.977,0.986,0.995 and 0.745,0.764,0.802 for predicting 1-,3-,and 5-year DFS in the training and validation sets,respectively.Multifactorial Cox regression analyses showed that clinical stage,carbohydrate antigen 125(CA125),and Radscore were the independent predictors of DFS.The area under the curve(AUC)of the combined model in the training and validation sets were 0.926,0.894,0.864 and 0.828,0.839,0.873 for predicting 1-,3-,and 5-year DFS.Meanwhile,bioinformatics analysis suggested that Radscore was significantly correlated with the immune level of endometrial cancer patients.Conclusions:Machine learning-based radiomics model is helpful for the prediction of DFS and immune levels in endometrial cancer patients.The combination of radiomics and clinical indicators can further improve the accuracy of prediction and provide a reference basis for prognostic prediction and individualized treatment of endometrial cancer patients.
作者
陈树清
张羽
陈东
常喜豹
刘静静
陈雷
钱银锋
CHEN Shuqing;ZHANG Yu;CHEN Dong;CHANG Xibao;LIU Jingjing;CHEN Lei;QIAN Yinfeng(Department of Imaging,Funan Hospital Affiliated Fuyang Normal University Medical College,Fuyang 236300,China;Department of Imaging,the First Affiliated Hospital,University of Science and Technology of China,Hefei 230001,China;Department of magnetic resonance,the First Affiliated Hospital of Anhui Medical University,Hefei 230022,China)
出处
《磁共振成像》
CAS
CSCD
北大核心
2024年第9期107-113,共7页
Chinese Journal of Magnetic Resonance Imaging
关键词
子宫内膜癌
磁共振成像
影像组学
机器学习
无病生存期
endometrial cancer
magnetic resonance imaging
radiomics
machine learning
disease-free survival
