毕业生知识库

  现代功能磁共振成像技术的产生为探索和揭示大脑的认知功能提供了新的途径。其中，静息态功能磁共振成像技术不需要研究者设计实验任务，不需要受试者执行实验任务，特别是对神经精神类疾病的患者来说，是一种更容易接受、更容易成功的技术。由于原始的功能磁共振信号包含了大量的信息，除了与神经活动密切相关的信号外，还有由其他生理因素如呼吸、心跳等引起的寄生信号和仪器产生的噪声，虽然预处理可以降低一部分影响，但是却难以消除。因此需要在研究中有效的从混杂信号中提取可用信息，并根据这些信息发现符合神经生理学意义的结果。在这一目的驱动下，出现了多种基于图像处理和统计方法的分析技术，大致分为两类：单变量方法和多变量方法。本文在前人提出的多变量分析方法基础上进行了深入的研究，优化了算法，取得了一些研究成果。此外，将提出的多变量分析方法应用到三个不同的研究中，一方面肯定分析方法的可行性，一方面为临床医学特别是临床诊断的相关研究提供新思路。本论文的具体研究内容和创新之处如下：

  1、改进的多变量分析方法。

  本论文结合了组信息指导下的独立成分分析方法，使独立成分具有更强的独立性和空间对应性。 首先基于信息论准则估计独立成分个数，利用组信息指导下的独立成分分析方法得到脑功能网络，并结合模式识别的分类方法和前向选择方法选取最具表达性的脑功能网络子集构成最优功能连接模式。

  其次，本论文改进了分析方法中支持向量机的输出值，改用输出后验概率值作为预测指标，一方面考虑了原有方法中决策值在各个二分类器中没有可比性，仅可以作为分类结果置信度的评价，另一方面也方便改进后的方法推广演绎在多目标分类器中。

  为了适应越来越多、越来越精细的脑网络，还改进了前向选择算法，提升算法效率，来帮助分析研究更细致的脑功能区域，而不至于付出巨大的计算代价。同时多层嵌套交叉验证过程使简化的前向选择方法得到的结果更加可靠、鲁棒。

  最后，本论文总结了提出的多变量分析方法流程与步骤，方便应用于不同的研究领域内。我们完成了代码的整合，使其泛化性能更好，多变量判别分析方法和组信息指导下的独立成分分析方法的源代码封装包将随着相关文章的发表而发布在https://www.nitrc.org/。

  2、多变量分析方法在神经精神类疾病研究中的应用

  多变量分析方法在磁共振影像分析中还处于发展阶段，很多新领域的研究都是以传统单变量分析为主，而本论文中的方法也还没有广泛的应用到相关领域中，本论文在不同背景和目的的静息态功能磁共振影像数据中运用本文提出的方法，一方面肯定分析方法的可行性，一方面为临床医学特别是临床诊断的相关研究提供新思路。

  精神分裂症是一种严重的、没有明确病因却具有较高遗传性的精神疾病，临床数据显示电休克疗法可以帮助精神药物治疗无效或对药物治疗不能耐受者的治疗。然而，电休克疗法对脑网络的改善作用与临床反应有怎样的相关性还未可知。利用本论文提出的多变量分析方法描述与精神分裂症患者相关的异常脑网络，发现其对精神分裂症患者的治疗结果具有一定的预测性，有助于临床医生识别出最可能从电休克疗法中获益的精神分裂症患者。

  此外，由于早期干预可能有助于预防和延迟精神分裂症的发作，因此我们希望找到生物标记物，能够在早期识别出精神分裂症高危人群。本论文通过多变量判别分析方法识别出异常的大尺度脑功能网络，它们的异常表现不仅存在于精神分裂症患者当中，而且在一级亲属中也发现了类似的情况。基于最优判别脑网络预测的具有精神分裂症特异性一级亲属在认知功能方面与精神分裂症患者相似，一级亲属和精神分裂症患者的分类特征值与数字符号编码分数相关。这些发现有助于从未发病的一级亲属中识别出具有精神分裂症特异性功能连接模式的人群，并可能帮助早期识别和治疗精神分裂症高危人群。

  最后，论文中还研究了一种罕见的可治愈的神经遗传疾病威尔逊氏病。由于其罕见，开展的研究相对较少，威尔逊氏病改变大尺度脑功能网络的机制与途径在很大程度上还是未知的。采用本论文提出的多变量分析方法，识别出与威尔逊氏病相关的脑网络，发现异常的功能网络对威尔逊氏病患者具有预测性，并可能帮助威尔逊氏病的早期识别和前期治疗，为其临床干预开辟新的途径。

  The modern functional magnetic resonance imaging (fMRI) techniques provide a new approach to explore and reveal the cognitive functions of the brain. Among them, the resting-state fMRI technique with no explicit stimuli does not require researchers to design experimental tasks, as well as participants to perform these experimental tasks. Particularly, some participants are with neuropsychiatric disorders that might have difficulty in carrying out the tasks. The major fluctuations of no interest that corrupt fMRI signal include rapid and slow head movements, physiological activity (breathing and heartbeat) and possible acquisition artifacts, which should be removed from the data to achieve structured noise reduction. Although preprocessing procedure can improve the signal-to-noise ratio (SNR) of fMRI signals, it is difficult to eliminate noise. Therefore, it is necessary to extract the available information signals effectively and to improve any subsequent detection and analysis of signal fluctuations related to neural activity. An increasing number of analyses based on image processing and statistical methods are proposed, which can be roughly divided into two categories: univariate methods and multivariate methods. In this paper, we have deeply studied the multivariate pattern classification techniques based on previous studies using resting-state, and achieved some progress for improving the performance of multivariate method. In addition, the proposed method was adopted to three different studies to verify the feasibility and superiority. Meanwhile, it may provide a new idea for the related research of clinical medicine, especially clinical diagnosis. In this paper, studies include the following parts:

  1、Modified multivariate analysis.

  A group information-guided independent component analysis (GIG-ICA) technique was used in this study to extract the subject-specific independent components (ICs) as functional networks (FNs) with stronger independence and better spatial correspondence across subjects. The number of ICs was estimated automatically based on information theoretic criteria. Then, a multivariate pattern classification method using forward component selection technique on Grassmann manifold was adopted to identify informative and discriminative FNs.

  Decision values of support vector machine (SVM) as confidence level of the evaluation result are not comparable between two different classifiers. Thus, instead of decision values as the output in previous method, posteriori probability values were as the output predictors. The improvement also can be extended for multi-class classification effectively in the future.

  To explore the subtle functional alteration based on more and more brain networks, a simplified forward component selection technique was proposed to select FNs for constructing the most discriminative set of FNs based Riemannian distance on the Grassmann manifold. It contributed to exploring more complex brain regions without huge computational costs. Additionally, to avoid any bias in the experiments, the models were built with a multi-level nested cross validation procedure to optimize the performance.

  Finally, the steps of the proposed multivariate analysis were summarized, which could be convenient to be applied in other researches. Source codes of the classification method and GIG-ICA will be available at https://www.nitrc.org/.

  2、The applications of multivariate analysis on neuropsychiatric disorders

  Multivariate analysis on fMRI is still in the development. Most of fMRI time series analyses are based on single voxel data evaluation using parametric statistical tests, especially in the new research fields of neuropsychiatric disorders, and our method has not been widely applied to related studies. In this paper, our proposed method was adopted to three different studies to verify the feasibility and superiority. Meanwhile, it may provide a new idea for the related research of clinical medicine, especially clinical diagnosis.

  Schizophrenia is a severe, disabling, and highly heritable psychiatric disorder with an unknown etiology. Electroconvulsive therapy is an effective treatment for schizophrenia, and it may serve as an augmentation strategy for treatment-resistant schizophrenia. However, remaining unclear is the way in which these ECT-induced changes in brain networks are related to the clinical response. This study adopted multivariate pattern recognition methods to characterize brain network patterns of schizophrenia patients and investigated relationship between brain network changes and symptomatic improvements in schizophrenia patients who received treatments of antipsychotics and a combination of antipsychotics and ECT. The results suggested that functional connectivity patterns are predictive for therapeutic outcomes in schizophrenia patients, andthese findings may be helpful for clinicians to identify patients who are most likely to benefit from ECT.

  Since early intervention could potentially help prevent and delay onset of psychosis, it is desired to develop biomarkers capable of early identifying individuals with higher risks of developing schizophrenia. The multivariate pattern classification method was used to learn informative large scale FNs. These aberrant FNs are informative for quantifying brain alternation in the unaffected first degree relatives (FDRs), not only in schizophrenia patients (SCZs). The SCZ-specific FDRs were similar to the SCZs, but significantly different from healthy controls (HCs) and HC-specific FDRs in terms of their cognitive measures. The classification scores of the FDRs and SCZs were both correlated with their cognitive measures. These findings suggested that pattern recognition of large scale FNs could help identify unaffected FDRs with schizophrenia-specific FC patterns.

  Finally, We also studied on a rare and treatable neurogenetic disease-Wilson’s disease (WD). Although the brain structural abnormalities associated with WD have been documented, it remains largely unknown how WD affects large-scale functional brain networks. This is the first study to investigate the large-scale FNs utilizing themultivariate pattern recognition technique in neurological WD patients. The present results suggested that aberrant FNs were predictive in neurological WD patients, and might help open up novel avenues for clinical interventions for neurological WD.