基于多模态MR影像的胶质母细胞瘤高危区域定位及预后预测研究

Heterogeneity is the important feature of glioblastoma multiforme (GBM). The heterogeneity of GBM contributes to its poor and variant prognosis. Due to lack of effective method to evaluate heterogeneity clinically, the radiomics may provide a new way to investigate it. In our previous radiomics studies, the texture features of magnetic resonance (MR) images have been used to evaluate the GBM heterogeneity. Meanwhile, we found that the prognosis information may be different in different region of GBM. However, in most radiomics approaches, the whole region of tumor was used. It is hard to evaluate the prognosis characteristics of different region in GBM. This may induce that the poor prognosis information of the high-risk regions in GBM was suppressed by the information of the low-risk regions, which may influence the accuracy of prediction. In this study, we have proposed a method to evaluate the GBM heterogeneity and establish the prognostic prediction model, based on multi-modality MR imaging. This approach mainly includes three parts. 1) Based on hierarchical clustering, the parenchyma of tumor is segmented into multiple subregions. 2) The prognosis is used to evaluate the dangerous level of each subregion. Then, the multiple instance learning method is used to identify the high-risk subregions. 3) A machine learning method is used to select the imaging biomarkers to represent the prognostic characteristics of subregions. The proposed method may combine the heterogeneity analysis and the prognostic prediction effectively and establish the prognostic prediction model. The results may have vital clinical significance, which would provide quantitative imaging information for the diagnosis of GBM and operation orientation.

异质性是造成胶质母细胞瘤(GBM)预后差且差异大的重要因素。目前的临床手段很难评价异质性，影像组学的提出为异质性研究提供了方法。申请人研究发现，MR影像的纹理特征可评价GBM异质性，且GBM内不同区域包含不同的预后信息。但目前的研究大都针对整个肿瘤区域，很难评价肿瘤内不同区域的预后特性，容易造成表征不良预后的“高危”区域信息被“低危”区域信息淹没，进而影响预后预测的准确性。为此，本项目提出了基于多模态MR影像的异质性分析及预后预测方法，主要包括：1）基于层次聚类方法，将GBM内具有异质性的肿瘤实质区域分割为多个同质子区域；2）将预后作为子区域危险程度的评价标准，利用多事件学习方法，判断各个子区域的危险特性，定位高危区域，实现异质性分析；3）筛选出可反映子区域预后特性的影像学标记，建立基于异质性分析的GBM预后预测模型。研究结果可为GBM诊断及手术定位等方面提供定量信息，具有重要临床意义。

异质性是造成胶质母细胞瘤(Glioblastoma, GBM)预后差且差异大的重要因素。目前的临床手段很难评价异质性，影像组学的提出为异质性研究提供了方法。申请人研究发现，MR影像的纹理特征可评价GBM异质性，且GBM内不同区域包含了不同的预后信息。但目前的研究大都针对整个肿瘤区域，很难评价肿瘤内不同区域的预后特性，容易造成表征不良预后的“高危”区域信息被“低危”区域信息淹没，进而影响预后预测的准确性。本项目紧扣“GBM异质性分析”和“GBM预后预测”两大临床问题，以GBM患者的多模态MR影像作为数据基础，开展了以下三方面工作。1）针对具有不同MR影像征象的子区域与GBM患者生存预后关系这一问题，将多模MR影像和GBM的多个区域结合，利用GBM患者的四种MR模态（T1增强、T1加权、T2加权和FLAIR序列）影像数据，提取了四个区域（整个肿瘤区域、增强区域、坏死区域和非增强/水肿区域）的影像特征，构建了基于多模态MR影像征象子区域的GBM患者生存分层诺模图。2）针对GBM瘤内是否存在与不良预后相关的高危子区域，以及这些高危子区域是否能够更好地指导患者生存分层这一问题，本项目提出了基于多模态影像的肿瘤异质性分割和基于多示例学习的弱监督学习算法，解决以患者级粗标记推断子区域标记的医学影像弱标记问题，并构建了基于多示例学习算法的生存相关高危子区域挖掘与生存长短分层模型；3）针对目前深度学习模型可视化定位不准确的问题，本项目提出了一个可兼顾肿瘤分层预测和结果可视化的SHNet (Subregion-based Hybrid Network)模型，可有效解决可视化分辨率低、注意力漂移等问题，在保证可视化展示区域准确的前提下对胶质瘤分级进行预测。通过本项目的研究，建立了肿瘤异质性分割、高危区域定位和深度可视化方法，可为GBM诊断及手术定位等方面提供定量信息，具有重要临床意义。