CASIA OpenIR  > 毕业生  > 博士学位论文
基于多模态脑影像的个体指标预测方法及其临床应用
姜荣涛
Subtype博士
Thesis Advisor隋婧
2020-05-27
Degree Grantor中国科学院大学
Place of Conferral中国科学院自动化研究所
Degree Name工学博士
Degree Discipline模式识别与智能系统
Keyword磁共振影像 脑功能连接 个体化预测 抑郁症 回归分析
Abstract

      随着机器学习算法的发展,神经科学的研究进入到了以多变量个体预测方法为主要手段的转化神经科学阶段。此类研究致力于利用先进的模式回归算法,开发、应用能够在个体水平对连续变量进行准确预测的多元分析模型,寻找稳健可靠的影像学标记物,并最终在教育和临床中发挥积极作用。目前,基于机器学习算法的预测模型在多种行为指标的个体差异检测中显示了巨大潜力。尽管如此,这些研究大都局限于使用单一模态的影像特征在小样本上建立模型,并且缺少独立中心数据对模型的泛化性能进行检验。此外,现有研究大都使用静息态功能磁共振影像(fMRI)提取的功能连接为输入特征,任务态功能连接对个体指标的预测潜力未被充分挖掘。基于此,本文主要进行了以下创新性工作:

      [1] 提出了一种包含体素级特征搜索、自适应空间聚类及模式回归的预测模型,利用基线状态的灰质密度特征对重度抑郁症患者接受电痉挛疗法后的恢复情况进行了定量预测,突破了现有研究基于二元分类方法的局限,并在三个独立中心数据上验证了模型的泛化性能,识别出了六个可作为预测电痉挛疗法疗效的潜在影像标记物,为临床开展具有针对性的治疗及预后提供了客观稳定的早期影像学靶点。此外,综合对比了单变量分析和多元预测模型的分析结果,说明具有预测性能的脑区不一定是治疗响应最大的脑区。该研究揭示了基于机器学习的影像标记物在个体化疗效预测方面巨大的临床转化价值,对辅助临床医师制定个性化医疗决策具有重要意义。

      [2] 提出了一种包含特征选择和稀疏回归的预测模型,对360名健康成人的智商评分进行了预测。本研究首次在多中心数据集上建立了具有优良泛化性能的智商预测模型,突破了现有研究缺少独立数据对模型进行验证的局限,并识别了若干条可作为智商潜在影像标记物的功能连接特征。研究创新性地从性别角度入手,探究了智商预测模型的性别特异性。结果发现,从男性样本中得到的功能连接只对男性的智力水平及认知子域具有预测性和相关性,对女性不具有预测性和相关性,反之亦然。此外,功能连接对女性智商具有更高的预测性能,这可能与男性智商受更加复杂的脑机制调控有关。该研究从个体预测的角度揭示了智商的产生不仅依赖以额顶网络为核心的高级认知系统,同样依赖基底神经节等皮下脑区,并且男女智商可能受不同神经机制所调控,为其他研究进行个体智商预测提供了重要参考。

      [3] 针对大部分智商预测研究仅使用单一模态特征的问题,本研究联合使用灰质皮层厚度和功能连接建立了多模态智商预测模型,取得了显著高于任何单一模态的预测精度,说明不同模态特征间存在对检测个体差异有益的互补信息。此外,本研究首创性地从男女各自认知子域优势的角度,对智商预测模型识别的多模态特征进行了解释。结果发现,对男性智商具有预测性能的多模态特征更多地分布在与空间推理、数学运算相关的脑区,而对女性智商具有预测性能的多模态特征集中分布在与语言处理能力相关的脑区,这与男女各自在认知领域上的优势是一致的。本工作充分说明了集成多模态特征能够从不同的测量视角揭示个体差异背后的脑功能结构基础,为未来其它研究从多模态的角度认识大脑提供了新思路。

      [4] 现有个体预测研究大都使用静息态功能连接,本工作基于463名健康成人样本,分别利用提取自静息态和7种任务态fMRI的功能连接特征,探究了任务态功能连接在个体差异研究中的潜力。结果发现,基于任务态功能连接的模型具有更高的预测性能,并且组合多种fMRI状态下的连接特征能够显著地提高认知预测精度,说明不同任务态数据中包含状态特异的个体差异信息。通过对多种个体指标进行分析,并充分考虑头动、影像扫描时间、交叉验证策略等一系列因素的影响,说明了该结论的稳定性和普适性。本研究首次在一个较大的数据集上,从个体化预测的角度系统地说明了大脑受内在网络架构调节,处于相对稳定的状态,认知任务的执行能够放大与所执行任务相关的神经环路中的个体差异信号,使大脑呈现出一些静息状态下不能检测到的个体差异信息,从而增强其对个体指标的预测能力。该研究证明了任务态fMRI是研究“脑-行为”关系的有力工具,为研究认知功能及精神疾病的病理机制提供了新依据。

Other Abstract

    With the rapid development of machine learning, the neuroimaging community is moving towards a translational neuroscience era, which is characterized by the use of multivariate predictive modeling. These studies maintain a focus on decoding individual differences in a continuously behavioral phenotype from neuroimaging data using regression-based methods, with an ultimate goal of identifying reliable neuromarkers that can aid in clinical practice. The machine learning-based predictive modeling has been successfully applied in the prediction of multiple important behavioral aspects. Despite such progress, most investigations either used limited sample sizes or lack of replication in independent cohorts. Additionally, resting state functional connectivity (FC) has been playing a dominating role in the field of predictive modeling, with the potential of task-induced FCs in individualized prediction largely unexplored. The main achievements of this paper are as follow:

    [1] A prediction model incorporating voxel-level feature selection, spatial clustering and pattern regression was proposed to predict the post-ECT (electroconvulsive therapy) remission status using pre-ECT grey matter for patients with major depressive disorders, which is different from previous studies using classification algorithm. The model achieved high accuracy for remission prediction in three independent data sets. Moreover, six grey matter regions were identified as predictive neuromarkers of ECT, providing stable and objective imaging target for clinical treatment and prognosis. Additionally, results derived from univariate brain mapping and multivariate predictive modeling demonstrated that the treatment-responsive brain regions may not be the most predictive ones. This study revealed the great value of developing machine learning-based neuromarkers in translational neuroscience and may provide potential opportunities for more effective and timely interventions in clinical.

    [2] Based on whole-brain FCs, a prediction framework integrating feature selection and sparse regression was proposed to predict the IQ scores for 360 healthy subjects. This study established a highly generalizable IQ-predictive model on a multi-center data set, breaking through the limitations of existing studies that lack replication in independent cohorts. Several predictive FCs that can be used as potential neuromarkers of IQ were identified. Moreover, this study investigated the gender specificity of IQ-predictive models, with the identified FC patterns uniquely predictive on and correlated with IQ and its sub-domain scores only within the same gender, but not for the opposite gender. Moreover, females exhibited significantly higher IQ-predictability than males, which can be partly attributed to a more complex substrate of intelligence in males. This study facilitated our understanding of the biological basis of intelligence by demonstrating that intelligence was underpinned by a variety of complex neural mechanisms that engage an interacting network of regions—particularly prefrontal-parietal and basal ganglia— whereas the network pattern differed between genders, providing an important reference for other studies.

    [3] Aiming at the limitation of most investigations using single-modal features of sMRI or fMRI for individual cognitive prediction, this study established a multimodal IQ-predictive model by integrating resting-state FC and grey matter cortical thickness. Results demonstrated that multimodal data achieved improved prediction accuracy than using any single modality alone, by capitalizing on complementary information provided by each single modality. Moreover, this study discussed the neurobiological substrates of gender differences underlying intelligence from the perspective of their respective superiority in cognitive domains. Specifically, females IQ demonstrated closer correlations with brain regions and functional network that are implicated in verbal learning and item memory; while, males IQ demonstrated closer correlations with brain cortex and functional network that are implicated in spatial cognition and logical thinking, which were consistent with their respective superiority in cognitive domain. This study highlighted the potential of using multimodal data to investigate the neural mechanizes of individual differences in behavioral phenotypes, providing new guidance on other studies to characterize the human brain.

    [4] Most existing studies in individualized prediction are commonly limited to using resting-state FCs. Using FCs derived from resting and 7 distinct task fMRI states for 463 healthy subjects, this study investigated the potential of task-induced FCs in probing individual differences. Results demonstrated that task-based machine learning models often outperformed rest-based models in phenotype prediction, and combining multi-task fMRI further improved prediction performance, signifying that individual differences in task-based FCs consist of state-dependent aspects. The reliability and generality of the above conclusion was confirmed by predicting multiple cognitive measures and conducting a series of control analysis including head motion, scan duration, and cross-validation scheme. Based on a relatively large cohort, this study indicated that the brain's large-scale networks were dominated by a stable intrinsic architecture, and tasks could perturb functional connections in the brain and further amplify individual differences in the neural circuitry underlying related traits, which can not be detected by resting state. Together, this study highlighted the benefit of using task-based FCs to reveal brain-behavior relationships, providing a new basis for the study of cognitive functions and the pathological mechanism of mental illness.

Pages130
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/39281
Collection毕业生_博士学位论文
Corresponding Author姜荣涛
Recommended Citation
GB/T 7714
姜荣涛. 基于多模态脑影像的个体指标预测方法及其临床应用[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
毕业论文.pdf(10953KB)学位论文 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[姜荣涛]'s Articles
Baidu academic
Similar articles in Baidu academic
[姜荣涛]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[姜荣涛]'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.