Value of combining radiomics and deep-learning with hematological inflammatory markers in predicting the prognosis of glioma
赵杉 1阎子康 2杨骏骏 3张文韬 1潘世娇 1徐胜生1
扫码查看
点击上方二维码区域，可以放大扫码查看
作者信息
1. 重庆医科大学附属第一医院放射科,重庆 400016
2. 重庆医科大学基础医学院生物信息系,重庆 400016
3. 重庆大学物理学院,重庆 400044
折叠
摘要
目的 探讨基于影像组学和深度学习特征评分(radiomics-deep learning score,RD-score)联合血液炎性指标构建的列线图在术前预测胶质瘤预后的价值.材料与方法 回顾性分析166例临床确诊胶质瘤的患者病例,按8:2随机分为训练集(133例)和验证集(33例).收集患者的临床、血液炎性指标资料,构建组合变量系统性炎症指数(systemic immune inflammation index,SII)、全身炎症反应指数(system inflammation response index,SIRI)、衍生中性粒细胞与淋巴细胞比值(derived neutrophil-to-lymphocyte ratio,dNLR)、中性粒细胞与淋巴细胞比值(neutrophil-to-lymphocyte ratio,NLR)、单核细胞与淋巴细胞比值(monocyte-to-lymphocyte ratio,MLR)、血小板与淋巴细胞比值(platelet-to-lymphocyte ratio,PLR)并计算其截断值.勾画胶质瘤感兴趣体积(volume of interest,VOI)并提取影像组学及深度学习特征,利用最小绝对值收敛和选择算子(least absolute shrinkage and selection operator,LASSO)-Cox进行特征筛选,构建基于影像组学特征评分(radiomics-score,Rad-score)、基于深度学习特征评分(deep learning-score,DL-score)以及RD-score模型,并比较三者的受试者工作特征曲线下面积(area under the curve,AUC)以评估其预测效能;基于RD-score对胶质瘤患者进行危险分层,并通过Kaplan-Meier生存分析绘制生存曲线.结合患者的临床因素、血液炎性指标和RD-score,使用多因素Cox回归构建术前预测总生存期(overall survival,OS)的RD-score模型、临床血液学模型和联合模型,计算AUC以评估各模型预测胶质瘤1、3、5年生存率的效能.绘制联合模型列线图,采用C指数(C-index)、校准曲线及决策曲线分析(decision curve analysis,DC A)评估列线图效能.结果 最终筛选出10个组学特征和8个深度学习特征用以构建RD-score.RD-score的预测效能高于Rad-score及DL-score(DeLong检验,P＜0.05),根据RD-score可将胶质瘤分为高风险组(RD-score≥1.09)和低风险组(RD-score＜1.09).多因素Cox回归结果显示年龄、肿瘤分级、术后化疗、SIRI和RD-score是胶质瘤预后的独立预测因素,基于以上因素构建的联合模型在训练集和验证集中的AUC高于RD-score模型及临床血液学模型(DeLong检验,P＜0.05).联合模型的可视化列线图预测OS的C-index分别为0.844和0.849;校准曲线提示在观察值和预测值之间有良好的一致性,DCA显示列线图有较高的净收益.结论 基于影像组学和深度学习的RD-score联合临床-血液炎性指标构建的列线图可以在术前有效预测胶质瘤患者的预后.
Abstract
Objective:To explore the value of a nomogram constructed by integrating radiomics and deep learning-based score(RD-score)with hematological inflammatory markers in preoperatively predicting the prognosis of glioma.Materials and Methods:A total of 166 clinically diagnosed glioma patients were retrospectively enrolled and randomly divided into a training set(133 cases)and a validation set(33 cases)in an 8:2 ratio.Clinical and hematological inflammatory marker of the patients were collected.Composite variables,including systemic immune inflammation index(SII),systemic immune response index(SIRI),derived neutrophil-to-lymphocyte ratio(dNLR),neutrophil-to-lymphocyte ratio(NLR),monocyte-to-lymphocyte ratio(MLR),and platelet-to-lymphocyte ratio(PLR)were constructed,and their optimal cut-off values were calculated.Delineating the volume of interest(VOI)for gliomas and extracting radiomics and deep learning features,utilizing least absolute shrinkage and selection operator(LASSO)-Cox for feature selection.Constructing radiomics-score(Rad-score),deep learning-score(DL-score),and RD-score,and comparing their receiver operating characteristic area under the curve(AUC)to assess predictive performance.Kaplan-Meier survival analysis was used to stratify glioma patients based on their RD-score.Integrating clinical data,hematological inflammatory marker,and RD-score,employing multivariable Cox regression to build RD-score model,clinical hematology model,and joint model to predict overall survival(OS).Calculating AUC to evaluate the efficiency of each model in predicting glioma 1,3,and 5-year survival rates.Drawing joint model nomogram and assessing their performance using C-index,calibration curves,and decision curve analysis(DCA).Results:After feature selection,10 radiomics features and 8 deep learning features were selected.The predictive performance of RD-score surpassed that of Rad-score and DL-score(DeLong test,P＜0.05).The constructed RD-score divided gliomas into high-risk group(RD-score≥ 1.09)and low-risk group(RD-score＜1.09).The results of the multivariable Cox regression showed that age,tumor grade,postoperative chemotherapy,SIRI,and RD-score were independent prognostic factors for glioma.The joint model,incorporating these factors,exhibited higher AUC in the training and validation sets compared to the RD-score model and clinical hematology model(DeLong test,P＜0.05).The visual nomogram of the joint model predicted OS with C-indices of 0.844 and 0.849 in the training and validation sets,respectively.Calibration curves indicated good consistency between observed and predicted values,and DCA demonstrated a high net benefit for the nomogram.Conclusions:The nomogram constructed by combining radiomics and deep learning-based RD-score with clinical-hematological inflammatory marker can effectively predict the prognosis of glioma patients preoperatively.
维普
万方数据