目的 观察基于MR-T2WI的深度迁移学习(deep transfer learning,DTL)特征、影像组学特征及临床特征构建的联合模型(列线图)在术前预测宫颈癌淋巴脉管间隙浸润(lymph vascular space invasion,LVSI)的价值.材料与方法 回顾性分析178例经术后病理证实为宫颈癌的患者病例,其中70例LVSI(+)、108例LVSI(-),按照8∶2划分为训练集[142例,54例LVSI(+)、88例LVSI(-)]和测试集[36例,16例LVSI(+)、20例LVSI(-)].对临床因素行单因素logistic分析,筛选出LVSI(+)独立预测因素.使用DTL方法和传统影像组学方法提取矢状位T2WI图像中病灶的DTL特征和影像组学特征,构建DTL特征数据集、影像组学特征数据集和DTL特征与影像组学特征融合的数据集,分别以t检验、Pearson分析和最小绝对收缩和选择算子(least absolute shrinkage and selection operator,LASSO)回归对训练集各特征数据集进行特征降维,以其最佳者构建影像组学(radiomics,Rad)模型、DTL模型、融合模型(Rad+DTL模型),并筛选最佳影像组学模型;基于上述最佳影像组学模型评分与临床独立因子构建联合模型,并绘制列线图.以校准曲线评估模型校准度,以决策曲线分析评价模型的应用价值.结果 淋巴结转移、粒细胞比率均为LVSI(+)的独立预测因子(P<0.05).Rad+DTL模型为最佳影像组学模型.联合模型在训练集和测试集中的受试者工作特征曲线下面积(area under the curve,AUC)高于Rad+DTL模型(0.984 vs.0.966,P<0.05;0.912 vs.0.759,P=0.05).联合模型的校准度较高,临床净收益更大.结论 基于MR-T2WI的DTL特征、影像组学特征联合临床特征构建的联合模型可有效预测宫颈癌LVSI.
Predicting lymph-vascular space invasion in cervical cancer based on MR-T2WI with deep learning and radiomic features combined with clinical features
Objective:To explore the value of preoperative prediction of cervical cancer lymph-vascular space invasion(LVSI)by combining deep transfer learning features based on MR-T2WI,radiomic features,and clinical characteristics.Materials and Methods:Data of 178 patients with cervical cancer by postoperative pathology,including 70 cases with LVSI(+)and 108 cases with LVSI(-)were retrospectively analyzed.The patients were divided into training set[n=142,including 54 LVSI(+)and 88 LVSI(-)]and test set[n=36,including 16 LVSI(+)and 20 LVSI(-)]at a ratio of 8∶2.Univariate logistic regression analysis was conducted on clinical factors to identify independent predictors for LVSI(+)cases.The deep transfer learning(DTL)method and traditional radiomics methods were used to extract the DTL features and radiomics features from the lesions in the sagittal T2WI images.This led to the construction of a deep transfer learning feature dataset,a radiomics feature dataset,and a dataset that merges the DTL features with the radiomics features.Each feature dataset in the training set underwent feature dimension reduction using t-tests,Pearson analysis,and least absolute shrinkage and selection operator(LASSO)regression.The best of these were used to construct radiomics models[radiomics(Rad)model,DTL model,Fusion model of Rad and DTL features(Rad+DTL)model],and the optimal radiomics model was selected.A joint model was constructed based on the best model's radiomics score and independent clinical factors,and a nomogram was drawn.The calibration of the model was evaluated using calibration curves,and the application value of the model was assessed using decision curve analysis.Results:Lymph node metastasis and the neutrophil-to-lymphocyte ratio were identified as independent predictors(P<0.05)with LVSI(+).The Rad+DTL model was determined as the optimal radiomics model.The combined model exhibited a higher AUC in the training set compared to the Rad+DTL model(0.984 vs.0.966,P<0.05),and in the testing set(0.912 vs.0.759,P=0.05).The combined model showed higher calibration accuracy and greater clinical net benefit.Conclusions:The combination of DTL features from MR-T2WI,radiomics features,and clinical characteristics can effectively predict LVSI in cervical cancer.
cervical cancerlymph-vascular space invasionradiomicsmagnetic resonance imagingdeep transfer learning
国家科技期刊平台
NSTL
万方数据
