基于多层感知机算法的CE-T_(1)WI影像组学模型预测胶质瘤MGMT启动子甲基化状态

A Multilayer Perceptron Algorithm-Based CE-T_(1)WI Radiomics Modelfor Predicting MGMT Promoter Methylation Status in Gliomas

目的 探索基于磁共振对比增强序列(CE-T_(1)WI)的影像组学模型预测脑胶质瘤患者O6-甲基鸟嘌呤-DNA甲基转移酶(MGMT)启动子甲基化状态的价值。方法 从公开数据集癌症影像档案库中回顾性分析281例脑胶质瘤患者(甲基化180例,未甲基化101例)的CE-T_(1)WI资料。最小绝对收缩和选择算子(LASSO)回归分析和最大最小冗余相关方法被用于特征筛选。采用多层感知机算法建立预测模型。受试者工作特征曲线下面积(AUC)被用于评价预测性能。结果 基于多层感知机算法的临床模型、影像组学模型以及临床+影像组学联合模型均可用于预测脑胶质瘤患者MGMT启动子甲基化状态,且与前两者相比,临床+影像组学联合模型有最高的诊断效能。联合模型在训练集中AUC为0.909(95%CI:0.864~0.944),敏感度和特异度分别为86.11%和82.50%;在验证集中AUC为0.831(95%CI:0.708~0.917),敏感度和特异度分别为80.56%和80.95%。Delong检验显示,临床+影像组学联合模型与临床模型之间AUC的差异具有统计学意义(训练集:Z=4.718,P<0.05;验证集:Z=2.677,P<0.05)。结论 基于多层感知机算法的CE-T_(1)WI影像组学模型能够对MGMT启动子甲基化和MGMT启动子未甲基化的胶质瘤患者进行有效鉴别。

Objective Exploring the value of magnetic resonance contrast-enhanced T_(1)-weighted(CE-T_(1)WI)-based radiomics model for predicting O6-methylguanine-DNA methyltransferase(MGMT)promoter methylation status in patients with gliomas.Methods MRI data from 281 glioma patients(108 methylation and 101 unmethylation)were retrospectively analyzed from the publicly available dataset The Cancer Imaging Archive(TCIA).The least absolute shrinkage and selection operator(LASSO)regression analysis and the max-relevance and min-redundancy correlation method were used for feature selection.We established a prediction model based on the multi-layer perceptron algorithm.The area under the subject operating characteristic curve(AUC)was used to evaluate the predictive performance.Results The clinical model based on the multi-layer perceptron algorithm,the MRI model,and the combined clinical+MRI model could be used to predict the MGMT promoter methylation of patients with glioma,and the combined clinical+MRI model had the highest diagnostic efficacy compared with the former two,with an AUC of 0.909(95%CI:0.864-0.944),and the sensitivity and specificity of 86.11%and 82.50%,respectively in the training set.The validation set AUC was 0.831(95%CI:0.708-0.917),with sensitivity and specificity of 80.56%and 80.95%,respectively Delong's test showed that the difference in AUC between the combined clinical+MRI model and the clinical model was statistically significant(training set:Z=4.718,P<0.05,validation set:Z=2.677,P<0.05).Conclusion The multi-layer perceptron model based on MRI can effectively identify patients with MGMT promoter methylation and MGMT promoter unmethylation gliomas.