CASIA OpenIR  > 中国科学院分子影像重点实验室
基于定量磁共振影像组学的脑肿瘤无创诊断研究
张帅通
Subtype博士
Thesis Advisor田捷
2020-05-24
Degree Grantor中国科学院大学
Place of Conferral中国科学院自动化研究所
Degree Discipline计算机应用技术
Keyword磁共振影像(MRI) 影像组学 无创诊断 半监督学习 脑肿瘤
Abstract

  脑肿瘤的术前精确诊断对于患者后续治疗方案及随访方案的个性化制定具有重要的指导意义。目前,临床上脑肿瘤的诊断(病理类型诊断和分子标志物诊断)常常依赖于神经外科手术后肿瘤组织样本的病理组织学检查或基因测序和术前肿瘤的穿刺活检。病理组织学检查或基因测序具有滞后性,无法为脑肿瘤患者个体化治疗方案的制定提供有效依据;穿刺活检仅能获得一小部分肿瘤样本,由于肿瘤的异质性,这种检查手段提供的肿瘤组织病理学信息有限,且常常出现采样偏倚,无法为脑肿瘤进行精确的诊断。因此,亟需研发脑肿瘤的术前精确诊断新方法,以辅助医生进行临床诊断工作。
  磁共振成像(Magnetic Resonance Imaging, MRI)是脑肿瘤诊断分析中一种最为常用的工具,可在宏观层面给观察者提供肿瘤组织的全貌信息及瘤周组织的微环境信息,是无创地观察和分析脑部肿瘤的重要途径。近年来,影像组学技术迅速发展,且在肿瘤临床诊断和预后预测上取得了较好的进展。本文针对脑肿瘤术前精确诊断难的问题,围绕着“组织病理学诊断”和“分子标志物诊断”两方面开展了基于定量化 MRI 影像特征分析的脑肿瘤术前诊断研究,并从如何获取关键影像组学特征角度对影像组学技术进行改进:1)提出了基于集成策略的最大相关最小冗余特征筛选方法,以避免小样本数据中的样本偏差,提高特征筛选结果的稳定性;2)利用 MRI 多序列、多参数的特性,结合脑肿瘤本身和瘤周微环境的影像特征以增加模型输入端有效信息量,提出了多栖息地影像组学分析框架,以提高脑肿瘤术前诊断精度;3)提出半监督特征学习框架,提取特异性和解释性较强的影像学特征,以提高模型的可解释性。本文的主要工作及贡献如下:
  1、脑肿瘤组织病理学诊断:针对非功能型脑垂体腺瘤组织病理学亚型缺乏术前诊断方法这一关键临床问题,本文从提高特征筛选方法稳定性角度,提出了基于 E-MRMR 的特征筛选算法,以在特定样本子集中筛选到对全部患者有预测价值的 MRI 影像特征,从而提高非功能型脑垂体腺瘤亚型的诊断精度。在 E-MRMR 特征筛选算法中,本文采用随机分层抽样的采样策略进行子样本选取,在每个子样本集合中采用 MRMR 特征筛选算法对特征进行重要性排序,最后利用加权投票方法将重要性序列集成为最终的特征重要性排序结果。在非功能型脑垂体腺瘤亚型诊断问题中,随着训练样本的增加,E-MRMR 特征重要性排序中前 20 个特征的最大加权二部图匹配序列相似指数高达 0.85。同时,基于 E-MRMR 算法的影像特征构建的非功能型脑垂体腺瘤亚型诊断模型在天坛医院验证数据集上准确率为 0.811,比基于 MRMR 算法的影像特征构建的模型性能提升大于 0.06。更值得注意的是,该工作被美国 Moffitt 癌症研究中心教授 Robert Gillies 和美国哈佛医学院助理研究员 Hugo JWL Aerts 在影响因子为 244.59 的期刊《CA: A Cancer Journal for Clinicians》上发表的文章引用,并对该工作进行了评价,称该工作提出的算法可以用于非功能型脑垂体腺瘤亚型诊断。
  2、脑肿瘤分子标志物诊断:针对脑胶质瘤分子标志物异柠檬酸脱氢酶(isocitrate dehydrogenase, IDH)突变情况缺乏术前有效诊断方法这一关键临床问题,本文从增加模型输入端有效信息量的角度,提出了一种基于多栖息地影像组学分析框架以实现脑胶质瘤 IDH 突变情况的精确诊断。不同于常规影像组学分析框架,该分析框架除了考虑脑胶质瘤肿瘤实质区域特性,还对瘤周水肿区域进行了定量的特征提取,以增加脑胶质瘤表征信息获取通道,通过综合分析肿瘤实质区域和瘤周水肿区域的影像特征,进而构建脑胶质瘤 IDH 突变情况预测模型,实现脑胶质瘤 IDH 突变情况的无创诊断。在独立验证集上,基于本文提出的多栖息地影像组学框架的脑胶质瘤 IDH 诊断模型的受试者工作特征曲线面积(area under the receiver operating characteristic curve, AUC)为 0.900,比基于常规影像组学框架的脑胶质瘤 IDH 诊断模型性能提升大于 0.03,比基于临床风险因素构建的模型性能提升大于 0.09。
  3、脑肿瘤分子标志物诊断:针对脑胶质瘤多种关键分子标志物(IDH 突变情况、1p/19 双缺失情况)缺乏术前有效诊断方法这一关键临床问题,和传统影像组学特征的特异性和可解释性差这一关键技术问题,本文从提取强特异性和强可解释性特征的角度,提出了一种面向脑胶质瘤多序列 MRI 的半监督学习框架。该框架采用自编码器通过数据驱动的方式从公开数据集 BRATS2015(无标签)脑胶质瘤数据中学习肿瘤的稀疏表征,然后结合有标签数据构建脑胶质瘤多种分子标志物诊断模型。在该半监督学习框架中,本文提出了将可自适应剔除背景区域的全局池化算法,从而避免背景区域(非脑胶质瘤实质和水肿区域)对强特异性和可解释性影像特征及脑胶质瘤分子标志物术前诊断模型的影响。与影像组学中人工设计的特征相比,从脑肿瘤 MRI 影像中学习到的特征具有更好的特异性。另外,基于这些稀疏表征的 IDH 和 1p/19q 诊断模型的 AUC 为 0.829 和0.648。
  本文围绕脑肿瘤术前精准诊断(组织病理学诊断和分子标志物诊断)难的问题,从稳定特征筛选框架设计、融合多信息通道的多栖息地影像组学框架设计、强特异性和可解释性特征设计三个方面开展了研究,提出了适用于脑肿瘤诊断的影像组学预测模型。

Other Abstract

    Precise diagnosis of brain tumor is essential for individualized treatment decision and follow-up strategy making before surgery. In current routine clinical practice, the diagnosis of brain tumor relies on the histopathological examination or gene sequencing of the tumor tissue samples obtained from neurosurgery or the preoperative biopsy. Neurosurgery could not provide timely and effective evidence for individualized preoperative treatment decision making. While biopsy could only provide limited information from a small part of tumor tissue due to tumor heterogeneity and sampling bias. Therefore, it is ugent to develop a novel approach for the preoperative precise diagnosis of brain tumor to aid clinicians in treatment decision making.
    As a common tool for the diagnosis and analysis of brain tumor, Magnetic Resonance Imaging (MRI) could provide complete information on tumor phenotype and tumor microenvironment. In recent years, radiomics advances rapidly and has made great progress in the diagnosis and prognosis of brain tumor. Addressing on the preoperative precise diagnosis of brain tumor and focusing on the two aspects of “histopathological diagnosis” and “diagnosis of molecular biomakers”, this dissertation carried out diagnostic analysis of MRI features on brain tumor, and improved the radiomics technology from the perspective of how to obtain the key imaging features: (1) proposed ensemble strategy-based maximal relevance and minimal redundancy (E-MRMR) to avoid the sample bias in small data, which imporoved the robustness of feature selection results; (2) proposed multi-habitat radiomics framework, in which, radiomic features from both tumor region and peritumoral microenvironment were extracted to increase the effective information in the input of model; (3) proposed semi-supervised learning framework which could extract features with strong specificity and interpretability to improve the interpretability of predictive model. The main innovations and contributions of our study are as follows:
    1. Histopathological diagnosis of brain tumor: To predict the histopathological subtypes of nonfunctioning pituitary adenomas (NFPA) before surgery, this dissertation proposed E-MRMR feature selection algorithm to improve the stability of feature selection results and to select the features with predictive value for the whole dataset in the given specific subset, and thereby improved the accuracy for preoperative diagnosis of NFPA. In the E-MRMR, this dissertation carried out MRMR feature selection algorithm on different subsets generated by stratified random sampling approach and then ensembled these feature selection results using weighted voting method. In this dissertation, as the size of training set increased, the similarity index based on maximum weighted bipartite graph matching of feature selection results genenrated by the proposed E-MRMR achieved larger than 0.85. Besides, the predictive model for NFPA subtypes based on E-MRMR yielded an accuracy of 0.811 on the validation cohort, which was 0.06 higher than the model based on MRMR. What’s more, Professor Robert Gillies in Moffitt cancer institution and assistant research Hugo in the Havard Medical School cited and commented on our study in CA: A Cancer Journal for Clinicians (impact factor: 244.59), saying that our proposed algorithm could be used for the prediction of NFPA subtypes.
    2. Diagnosis of molecular biomarkers in brain tumor: To predict molecular biomarker IDH in gliomas preoperatively, this dissertation proposed multi-habitat-based radiomics framework which could improve the accuracy for prediction of IDH by increasing the effective information in the input of model. Different from traditional radiomics framework, the proposed framework extracted features from both tumor region and peritumoral microenviroment in account to increase the effective information for characterization of gliomas phenotype. In the validation cohort, the predictive model generated by the multi-habitat-based radiomics yielded an area under the receiver operating characteristic curve (AUC) of 0.900, which was 0.03 higher than the model by the traditional radiomics framework and 0.09 higher than the model based on clinical risk factors.
    3. Diagnosis of molecular biomarkers in brain tumor: To provide a predictive diagnosis model with strong specificity and interpretability for molecular biomarker IDH and 1p/19q in gliomas, this dissertation proposed a semi-supervised learning framework to extract strong specificity and interpretability imaging features from multi-sequence MRI. The proposed semi-supervised learning framework applied residual convolutional auto encoder to learn sparse representation of gliomas on the open dataset BRATS2015 (unlabelled) in a data-driven approach, and then developed predictive models for IDH and 1p/19q on the labelled dataset. In this semi-supervised learning framework, this dissertation proposed an adapative global pooling algorithm, which could screen out backgroup and thereby avoided the influence of background on sparse representation and models for the prediction of molecular biomarkers in gliomas. Compared with manually-designed radiomic features, these sparse representations were leart from MRI of gliomas, and thereby was of better specificity. In addition, predictive models for IDH and 1p/19q based on these sparse representations yielded AUC values of 0.829 and 0.648, respectively.
    Focusing on the precise diagnosis of brain tumors preoperatively, this dissertation proposed three methods including design of stable feature selection algorithm, design of multi-habitat-based radiomics framework by combining multi-channel information, and design of novel features with strong specificity and interpretability, and proposed corresponding radiomics models for diagnosis of brain tumors.

Pages124
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/39232
Collection中国科学院分子影像重点实验室
Recommended Citation
GB/T 7714
张帅通. 基于定量磁共振影像组学的脑肿瘤无创诊断研究[D]. 中国科学院自动化研究所. 中国科学院大学,2020.
Files in This Item:
File Name/Size DocType Version Access License
博士学位申请毕业论文.pdf(5591KB)学位论文 开放获取CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[张帅通]'s Articles
Baidu academic
Similar articles in Baidu academic
[张帅通]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[张帅通]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.