基于磁共振T2WI影像组学模型对胎盘植入性疾病进行产前诊断及分型

Radiomics model based on MR T2WI for prenatal diagnosis and classification of placenta accreta spectrum disorders

目的探讨基于磁共振T2WI的影像组学模型在产前预测胎盘植入性疾病(placenta accreta spectrum disorders,PAS)及其亚型的应用价值。材料与方法回顾性分析了2018年1月至2023年1月在北京妇产医院住院分娩的193例单胎妊娠孕妇数据,其中PAS 134例,非PAS 59例,所有患者根据同一分型的总数按2∶1的比例随机划分为训练集和测试集。在T2WI序列图像提取影像组学特征,Pearson相关系数和最小绝对收缩和选择算子(least absolute shrinkage and selection operator,LASSO)回归用于特征筛选,基于筛选后的特征构建PAS预测模型。然后,计算便于临床应用的影像组学评分评估PAS分型,使用单因素分析与多因素分析进一步分析其他潜在的临床危险因素,包括年龄、孕周、此前孕次、此前产次、此前剖宫产次数、胎盘问题(前置胎盘)和既往子宫手术史,选择临床主要风险因素建立基于影像组学评分和临床特征的临床-影像组学模型并绘制诺莫图。通过受试者工作特征(receiver operating characteristic,ROC)曲线下面积(area under the curve,AUC)评估模型的预测性能,采用DeLong检验比较模型间的预测效能,校准曲线用于评估预测模型的校准程度,决策曲线用于评估预测模型临床价值。结果在T2WI序列图像上提取了806个影像组学特征,经过Pearson相关分析后保留147个影像组学特征,经LASSO回归处理后筛选出10个影像组学特征,基于影像组学特征构建影像组学模型。影像组学模型的训练集AUC值为0.933(95%CI:0.888~0.978),准确率为88.37%,敏感度为88.78%,特异度为87.10%,阳性预测值(positive predictive value,PPV)为95.60%,阴性预测值(negative predictive value,NPV)为71.05%;测试集AUC值为0.914(95%CI:0.835~0.993),准确率为89.06%,敏感度为90.91%,特异度为85.00%,PPV为90.00%,NPV为80.95%。校准曲线和决策曲线表明模型具有较高性能和潜在临床应用价值。影像组学评分对穿透性胎盘植入具有较强的识别能力,训练集和测试集准确率分别为82.95%、89.06%,敏感度和NPV在训练集和测试集都达到了100.00%,特异度分别为81.35%、88.33%。同时,本研究成功构建了临床-影像组学模型并绘制诺莫图用于可视化预测患者的PAS,训练集中临床-影像组学模型的AUC值为0.969(95%CI:0.946~0.993),测试集中AUC值为0.976(95%CI:0.947~1.000)。DeLong检验测试结果表明两模型性能存在显著性差异(P<0.05),临床-影像组学模型具有更好的性能表现。结论基于临床特征及影像组学评分构建的临床-影像组学模型预测效能较好,可作为产前预测是否存在PAS的方法。且影像组学评分对PAS亚型具有较好的鉴别能力,尤其是对于穿透性胎盘植入。

Objective:To investigate the application value of radiomics model based on MR T2WI for prenatal predicting placenta accreta spectrum disorders(PAS)and determining the subtype of PAS.Materials and Methods:The data of 193 pregnant women with singleton pregnancies who were hospitalized for delivery in Beijing Obstetrics and Gynecology Hospital from January 2018 to January 2023 were retrospectively analyzed,including 134 cases of PAS and 59 cases of non-PAS.All pregnant women were randomly divided into training set and test set in a 2∶1 ratio based on the total number of patients with the same subtype.The radiomics features were extracted from T2WI sequence,Pearson correlation coefficient and least absolute shrinkage and selection operator(LASSO)regression were used for feature screening,and the radiomics models for predicting PAS were constructed.Then,a radiomics scoring system for clinical application is constructed and trained to evaluate the subtypes of PAS,and univariate analysis and multivariate analysis are used to further analyze other potential clinical risk factors,including age,gestational age,previous gravidity,previous parity,the history of cesarean section,placental problems(placenta previa),and the history of uterine-related operations.Establish a nomogram based on the selection of clinical major risk factors.The receiver operating characteristic(ROC)curve was drawn to evaluate the predictive performance of the model,and DeLong test was used to compare the predictive efficiency of these models,the calibration curve is used to evaluate the degree of calibration of the prediction model,and the decision curve is used to evaluate the clinical value of the prediction model.Results:806 radiomics features were extracted from T2WI sequence,147 radiomics features were retained after Pearson correlation analysis,and 10 radiomics features were screened out after LASSO regression processing,and a radiomics model that is applied to scoring was established.The area under the curve(AUC)value of the radiomics model in the training set was 0.933(95%CI:0.888-0.978),the accuracy was 88.37%,the sensitivity was 88.78%,the specificity was 87.10%,the positive predictive value(PPV)was 95.60%,and the negative predictive value(NPV)was 71.05%;the AUC value in the test set was 0.914(95%CI:0.835-0.993),the accuracy was 89.06%,the sensitivity was 90.91%,the specificity was 85.00%,the PPV was 90.00%,and the NPV was 80.95%.The calibration curve and decision curve showed that the model had high performance and potential clinical application value.The radiomics scoring model has a strong ability to identify placenta percreta,the accuracy of training set and test set reached 82.95%and 89.06%,the sensitivity and NPV reached 100.00%in training set and test set,and the specificity reached 81.35%and 88.33%,respectively.In addition,this study successfully constructed a clinical-radiomics model and draws a nomogram for visualizing PAS in patients.In the training set,the AUC of the clinical-imaging model was 0.969(95%CI:0.946-0.993),in the test set,the AUC value was 0.976(95%CI:0.947-1.000).DeLong test results showed that there were significant differences in the performance of the two models(P<0.05),and the clinical-imaging model had better performance.Conclusions:The clinical-radiomics model based on the clinical major risk factors and radiomics scoring system has a good performance,and can be used as a method to predict the presence of PAS before delivery.And the radiomics scoring system has a good ability to distinguish the subtype of PAS,especially the placenta percreta.