多序列临床-影像组学模型预测儿童髓母细胞瘤分子亚型

Multi-sequence clinical-radiomic model predicts molecular subgroups of pediatric medulloblastoma

目的探讨基于多序列磁共振成像(MRI)影像组学和临床特征的模型对儿童髓母细胞瘤(MB)分子亚型的预测价值。方法回顾性分析2011年1月至2020年1月郑州大学第一附属医院100例原发性MB患儿的MRI及临床资料,简单随机抽样选出50例患儿作为训练集,另外50例患儿作为测试集。训练集包括5例WNT激活型MB(WNT),5例SHH激活型MB(SHH),28例非WNT/SHH激活型Group3型MB(Group3),12例非WNT/SHH激活型Group4型MB(Group4),测试集包括11例WNT,3例SHH,24例Group3,12例Group4。从5929个在人工勾画的肿瘤区域提取的三维影像组学特征中选择稳健、非冗余的特征,进一步采用Boruta算法选择最优特征,将筛选出的影像组学特征基于训练集(50例)建立随机森林预测模型,并在测试集(50例)进行评估。结合影像组学特征与临床特征,建立随机森林联合预测模、临床-影像组学模型。结果包含13个最优影像组学特征的影像组学模型对分子亚型进行预测,在测试集中,WNT的受试者工作特征曲线下面积(AUC)值为0.9231,SHH、Group3、Group4的AUC值分别为0.6737、0.5192、0.7050,在加入临床特征后,测试集中WNT和SHH的AUC值提升至0.9441和0.8191。结论多序列临床-影像组学模型对儿童MB WNT和SHH分子亚型具有较高的预测价值,可以为患儿个性化诊疗提供决策支持。

Objective To explore the value of the model based on multi-sequence magnetic resonance imaging(MRI)radiomics and clinical features in predicting molecular subtypes of pediatric medulloblastoma(MB).Methods MRI imaging data and clinical data of 100 children with primary MB admitted in the First Affiliated Hospital of Zhengzhou University from January 2011 to January 2020 were analyzed retrospectively.Fifty children with primary MB were allocated to training cohort,and those of the other 50 were allocated to testing cohort by using simple random sampling method.In the training cohort,there were 5 cases of WNT-activated MB(Wingless,WNT),5 cases of SHH-activated MB(Sonic hedgehog,SHH),28 cases of non-WNT/non-SHH medulloblastoma Group3(Group3),12 cases of non-WNT/non-SHH medulloblastoma Group4(Group4).The testing cohort included 11 cases of WNT,3 cases of SHH,24 cases of Group3 and 12 cases of Group4.The robust and non-redundant features were selected from 5929 three-dimensional radiomic features extracted from the manually delineated tumor area,and Boruta algorithm was used to further select the optimal features.Based on the selected features,a random forest prediction model was constructed using the training cohort(50 cases),which was further used to evaluate the testing cohort(50 cases).Combined with radiomic features and clinical features,a joint random forest prediction,clinical-radiomic model was constructed.Results A radiomic model containing 13 optimal radiomics features was used to predict molecular subtypes of MB.The area under curve(AUC)of receiver operating characteristic(ROC)curve for WNT,SHH,Group3 and Group4 MB cases in the testing cohort was 0.9231,0.6737,0.5192 and 0.7050,respectively.Incorporating clinical features into the radiomic model improved AUC for WNT and SHH at 0.9441 and 0.8191,respectively.Conclusions The multi-sequence clinical radiomic model has a high predictive value for pediatric MB with the molecular subtypes of WNT and SHH,which provides decision-making supports for individualized diagnosis and treatment of pediatric MB.