基于多参数MRI的影像组学模型对子宫内膜样腺癌微卫星不稳定状态的术前预测价值

Multiparameter magnetic resonance imaging-based radiomics models for the preoperative prediction of microsatellite instability status in endometrioid adenocarcinoma

目的探讨基于多参数MRI影像组学模型术前预测子宫内膜样腺癌患者微卫星不稳定(MSI)状态的应用价值。方法回顾性队列研究。纳入2017年1月—2021年12月山西省肿瘤医院经病理证实为子宫内膜样腺癌的患者225例。患者年龄29~78(55.1±7.5)岁,均行全子宫+双附件切除术,并于术前行盆腔多参数MRI检查。将225例患者按7∶3的比例随机分为训练集(158例)和验证集(67例),根据手术标本病理免疫组织化学检查结果分为高度微卫星不稳定(MSI-H)组58例和低度微卫星不稳定/微卫星稳定(MSI-L/MSS)组167例。对每例患者的快速自旋回波压脂序列T2加权像、增强T1加权像、表观弥散系数(ADC)3个序列进行图像分割和特征提取,在训练集中采用组内相关系数(ICC)、Mann-WhitneyU检验、Spearman相关及Boruta算法进行图像特征筛选,并使用朴素贝叶斯(NB)、随机森林(RF)及支持向量机(SVM)3种分类器构建影像组学模型。在训练集、验证集中分别采用受试者工作特征(ROC)曲线、Delong检验、决策分析曲线(DCA)评估并验证3种模型对MSI状态的诊断性能、预测性能、净收益。将筛选出的影像组学特征纳入诊断效能较高的最佳模型中,对每例患者进行影像组学评分,分别在训练集、验证集中评估MSI-H组和MSI-L/MSS组影像组学评分的分布情况。结果训练集与验证集间比较,患者肿瘤分化程度差异有统计学意义(χ^(2)=8.40,P=0.015),患者年龄、体质量指数、国际妇产科联盟分期,以及肌层浸润深度等临床病理资料差异均无统计学意义(P值均>0.05)。在训练集中,MSI-H组与MSI-L/MSS组比较,患者是否绝经2组间差异有统计学意义(χ^(2)=4.56,P=0.033),其他临床病理资料2组间差异均无统计学意义(P值均>0.05);在验证集中,2组间各临床病理差异均无统计学意义(P值均>0.05)。多参数MRI图像经特征提取后,每例患者获得4245个影像组学特征,在训练集中筛选出6个关键特征构建模型。在训练集中NB模型、RF模型及SVM模型预测子宫内膜样腺癌患者MSI状态的ROC曲线下面积(AUC)分别为0.764[95%可信区间(CI)0.682~0.846]、0.821(95%CI 0.751~0.892)和0.905(95%CI 0.848~0.961),验证集中AUC分别为0.712(95%CI 0.568~0.856)、0.812(95%CI 0.710~0.915)、0.875(95%CI 0.762~0.988),结果表明SVM模型的诊断性能最佳。Delong检验结果显示,训练集中SVM模型与NB模型、RF模型AUC间差异均有统计学意义(Z=-3.45、-2.33,P值均<0.05),而在验证集中AUC差异均无统计学意义(P值均>0.05)。DCA曲线显示,在训练集和验证集中,SVM模型预测MSI状态的净收益均优于NB模型、RF模型,也优于将所有患者MSI状态都看作MSI-H或MSI-L/MSS。采用最佳模型SVM模型对每例患者进行影像组学评分,训练集及验证集中MSI-H组与MSI-L/MSS组影像组学评分分布有明显差异。结论基于多参数MRI的SVM模型对于子宫内膜样腺癌患者MSI状态具有一定的术前预测价值。

Objective This study aimed to evaluate the application value of multiparameter magnetic resonance imaging(MRI)-based radiomics models for the preoperative prediction of microsatellite instability(MSI)status in patients with endometrioid adenocarcinoma.Methods This retrospective cohort study included 225 patients,with an average age of 29-78(55.1±7.5)years,with pathologically proven endometrioid adenocarcinoma in Shanxi Cancer Hospital between January 2017 and December 2021.All patients underwent pelvic MRI scans before complete hysterectomy with oophorotomy and were randomly divided into training set(158 cases)and validation set(67 cases)at a ratio of 7∶3.Patients were divided into MSI-H group(58 cases)and MSI-L/MSS group(167 cases)according to pathological IHC results.Image segmentation and features were extracted on T2-weighted imaging fat suppression,contrast-enhanced T1WI images,and apparent diffusion coefficient(ADC)maps for each patient.In the training set,intraclass correlation coefficient(ICC),Mann-Whitney U test,Spearman correlation analysis,and Boruta algorithm were used for feature selection.Three classification algorithms including naive Bayesian algorithm(NB),random forest(RF),and support vector machine(SVM)were applied to build radiomics models.The diagnostic performance,predictive performance,net benifit,and reliability of the three radiomics models were tested by ROC curve,Delong test,and decision analysis curve(DCA)and verified in the validation set.The selected imaging features were included in the best model,and radscore measurement was performed for each patient.The distribution of radscore in the MSI-H and MSI-L/MSS groups was evaluated in the training and validation sets,respectively.Results Between the training and validation sets,the difference between patients'tumor grade was significant(χ^(2)=8.40,P=0.015)but not between other indicators such as age,body mass index,Federation of Gynecology and Obstetrics,and depth of myometrial invasion(all P values>0.05).In the training set,patients before or after menopause had statistically significant difference between the MSI-H and MSI-L/MSS groups(χ^(2)=4.56,P=0.033).The other indicators were not statistically different between the two groups(all P values>0.05).In the validation set,clinical and pathological data were not different between the two groups(all P values>0.05).After feature extraction of multiparameter MRI images,4245 imaging features were obtained for each patient.Six key radiomics features were selected for model building.The area under curve(AUC)of the NB,RF,and SVM models were 0.764(95%credibility interval[CI]0.682-0.846),0.821(95%CI 0.751-0.892),0.905(95%CI 0.848-0.961)and 0.712(95%CI 0.568-0.856)as well as 0.812(95%CI 0.710-0.915),0.875(95%CI 0.762-0.988)in the training and validation sets,respectively.The SVM model had the best performance in the training and validation sets.Delong test results showed that in the training set,the AUC difference in the SVM,NB,and RF models was statistically significant(Z=-3.45,-2.33,all P values<0.05),and that in the validation set was not statistically significant(all P values>0.05).The DCA results showed that the SVM model achieved higher net benefits than the other models and was better than treating all patients as MSI-H or MSI-L/MSS in the training and validation sets.The best model SVM was used for radscore measurement for each patient,and no significant differences in the distribution of radscore were found between the MSI-H and MSI-L/MSS groups in the training and validation sets.Conclusion The SVM model based on multi-parameter MRI can be used for the preoperative evaluation of MSI status in patients with endometrioid adenocarcinoma.