深度学习结合MR影像异质性的胶质母细胞瘤MGMT启动子甲基化预测研究

Accurate prediction of O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation is crucial to clinical decision-making of glioblastoma (GBM) patients. Currently, there is lack of effective method to detect MGMT promoter methylation before surgery, while studies indicate that multimodal magnetic resonance imaging (MRI) have the potential of noninvasive prediction of MGMT status. The preliminary studies confirm that the exploration of MR imaging heterogenous characteristics of GBM can help improve the prediction efficiency of MGMT. Moreover, the application of the deep convolutional neural network (CNN) model provides the feasibility of the cascade of the segmentation of GBM image heterogenous subregion and the prediction pf MGMT status. Focusing on the aim of predicting MGMT promoter methylation, we plan to construct a cascaded double-task deep convolutional neural network (CNN) model which has the ability of both MRI heterogeneous subregion segmentation and MGMT promoter methylation prediction. The model would be trained using transfer learning and layer-wise fine-tuning strategy, which can solve the CNN training problem caused by small sample image dataset. Then, to verify the superiority of CNN model, comparison of the prediction capacity of MGMT promoter methylation will be done between CNN-generative features and handcraft features that used in previous studies. This project is promising to propose a noninvasive technique for preoperative MGMT promoter methylation prediction and has important research and application value.

准确预测O6-甲基鸟嘌呤-DNA甲基转移酶(MGMT)启动子甲基化，对胶质母细胞瘤(GBM)患者的临床决策意义重大。目前缺乏术前检测MGMT启动子甲基化的有效手段，而研究表明多模态磁共振影像(MRI)具有无创预测MGMT状态的潜能。申请人前期研究证实，深入挖掘GBM的MR影像异质性子区域特性有助于提高MGMT状态预测效能；且深度卷积神经网络（CNN）模型的引入为GBM影像异质性区域分割与MGMT状态预测的级联提供了可能。由此本项目围绕“MGMT启动子甲基化预测”这一问题，拟建立具有多模态MR影像异质性子区域分割与MGMT启动子甲基化预测能力的级联式双任务CNN模型，并采用迁移学习与逐步微调策略解决小样本CNN模型训练难题；比较已建立CNN模型与常规影像组学模型的MGMT状态预测能力，验证前者的优越性。本项目有望为GBM患者提供一种MGMT术前无创预测手段，具有重要研究与应用价值。

准确预测O6-甲基鸟嘌呤-DNA甲基转移酶(MGMT)启动子甲基化，对胶质母细胞瘤(GBM)患者的临床决策意义重大。目前缺乏术前检测MGMT启动子甲基化的有效手段，而研究表明多模态磁共振影像(MRI)具有无创预测MGMT状态的潜能。申请人前期研究证实，深入挖掘GBM的MR影像异质性子区域特性有助于提高MGMT状态预测效能；且深度卷积神经网络（CNN）模型的引入为GBM影像异质性区域分割与MGMT状态预测的级联提供了可能。由此本项目围绕“MGMT启动子甲基化预测”这一问题，拟建立具有多模态MR影像异质性子区域分割与MGMT启动子甲基化预测能力的级联式双任务CNN模型，并采用迁移学习与逐步微调策略解决小样本CNN模型训练难题；比较已建立CNN模型与常规影像组学模型的MGMT状态预测能力，验证前者的优越性。本项目对磁共振图像异质性子区域分割，得到①r-EA：全瘤区；②r-CE：增强区；③r-Nec：坏死区④r-E/nEC：水肿区/非增强区四个区域，提取4000个纹理特征。利用弹性网络回归对特征进行降维筛选，计算λ值，得到前19个特征为最优特征子集。利用最优特征子集计算出特征系数Radiomics Signature ，建立诺莫图模型。通过输入患者的（年龄/性别/Radiomics Signature），计算标尺数值，得出该患者是MGMT甲基化的概率。矫正曲线和决策曲线分析说明模型具有良好的诊断效能和泛化性（矫正曲线的C-index在测试组和验证组分别为0.854和0.852）。本研究建立的术前无创预测胶质瘤患者MGMT启动子甲基化状态的模型能有效地预测患者的MGMT启动子甲基化状态，具有重要研究与应用价值。