脑网络组研究知识库

抑郁症是一种以显著持久的情绪低落为基本特征的情绪障碍类疾病，在临床症状和病程表现上具有较强的异质性，并与其他情绪障碍类疾病存在共同表征；由于其病因复杂、发病机制尚未明确，亟需客观稳定的影像学标记辅助其诊断和治疗。功能磁共振影像（fMRI）具有无创和空间分辨率高等特点，已在多项研究中被用于揭示抑郁症患者存在的异常脑功能活动和连接损伤，但在大脑功能的动态变化过程中，抑郁症患者是否存在特异性的动态连接模式，以及是否与特定的脑结构特征协同变化仍不清楚。且随着抑郁症大数据的不断发布，能否基于先进的数据驱动和深度学习算法，利用fMRI数据挖掘抑郁症临床诊断和亚型识别的影像学标记是当前研究的热点和难点，有待深入探索。由此启发，本文基于磁共振影像从大脑功能动态变化、脑功能结构协同共变、疾病分类和亚型识别四个方面深入研究了抑郁症患者，主要创新性工作归纳如下：

1．采用基于组信息的独立成分分析（GIG-ICA）和滑动时间窗方法，从全脑动态功能连接变化的视角系统地探索了抑郁症动态功能网络连接（dFNC）及其网络拓扑属性的异常。结果发现1) 大脑主要存在五种不同的动态功能连接模式。相比健康对照被试，重性抑郁症患者在和自我沉思相关的弱连接模式下停留更久的时间；2) 在全脑弱连接模式下，抑郁症患者的大脑网络全局效率、节点强度和节点中心性显著降低；3) 抑郁症患者在各状态下的异常功能连接主要与前额叶、小脑以及躯体运动网络相关，且不同状态下的共变连接和抑郁症患者的症状严重性、执行功能和空间记忆能力显著相关。本研究创新性地基于数据驱动算法探索了重性抑郁症全脑动态功能网络连接的异常，发现和抑郁症相关的全脑弱连接模式，对理解抑郁症脑功能网络异常的神经机制具有重要启发意义。

2．基于多变量典型相关分析联合动态功能连接和灰质体积特征，从多模态脑影像融合的角度探讨了抑郁症患者脑功能结构协同共变的关系，并在两个独立数据集上验证了抑郁症患者脑功能结构共变模式的稳定性和一致性。结果发现抑郁症患者脑功能结构的共变成分主要包括异常的额顶网络连接以及背外侧前额叶、海马、壳核、丘脑和颞下回下降的灰质体积。此外，相比静态功能连接，海马和丘脑下降的灰质体积与动态功能连接显著相关。本研究从全脑动态功能连接和灰质体积融合共变这一新的角度发现了背外侧前额叶和颞下回等脑区功能结构的异常，有助于对抑郁症功能结构协同共变的神经病理学基础更完整的理解。

3．基于最新的原型表征学习技术，结合大脑网络拓扑属性，提出了基于脑网络的卷积原型学习算法，能够自动为每个类别学习类间分离、类内紧凑的原型表征；同时结合saliency map的深度可解释性方法，实现对输入特征重要性的评估。将该模型应用于抑郁症分类研究发现: 1) 相比三个传统分类方法和两个深度学习网络，脑网络卷积原型学习算法对抑郁症的分类准确率提高2.4~7.2%；2) 对抑郁症分类贡献较大的功能连接主要位于皮下核团网络、额顶网络和默认网络，特别是位于前额叶皮层-皮下核团环路之间的连接。本研究提出了新的基于脑网络的卷积原型学习算法，提高了抑郁症的分类准确率并揭示了前额叶皮层-皮下核团环路在抑郁症分类中的重要性。

4．基于抑郁症大样本数据集（n=971），采用非负矩阵分解算法开展了抑郁症亚型识别，系统地研究了各亚型全脑功能连接的异质性，及其在临床电休克抗抑郁治疗中的疗效响应差异。结果发现1) 三种稳定复现的抑郁症亚型，其中额顶网络连接亚型和焦虑症状相关；2) 三种抑郁症亚型的跨数据集泛化性的平均分类准确率为89.2%；3) 更重要的是，额枕网络连接亚型在电休克抗抑郁治疗中表现出接近100%的响应率，显著高于其他两个抑郁症亚型（30%~70%），并在两个独立的电休克抗抑郁治疗数据集上得到验证。本研究基于大样本抑郁症数据集开展了基于数据驱动的亚型分析，发现了稳定复现的三种抑郁症亚型，表明了不同功能连接模式的抑郁症亚型可能表现出不同的临床症状和治疗响应，为理解抑郁症临床症状的个体差异性提供了新视角以及对推进抑郁症个体化临床诊疗具有重要指导意义。

Characterized by persistent anhedonia, depression is highly heterogeneous on clinical symptoms and longitudinal course of the illness, and has high comorbidity with other mental disorders. Due to its complex etiology and unclear pathogenesis, it is urgent to find objective and reliable imaging biomarkers for depression to assist its diagnosis and treatment. Functional magnetic resonance imaging (fMRI) provides a non-invasive method to investigate brain function with high spatial resolution, which has been widely used to detect and characterize brain networks, especially through functional connectome among spatially separated brain regions. However, in the dynamic changes of brain functional states, whether there are specific dynamic connectivity patterns, and covarying patterns for brain function and structure in depressive patients are still unclear. With the release of big data for depression, it is a hot topic and difficult problem to mine imaging biomarkers for the diagnosis and subtype recognition for depression. Motivated by this, we explore the aberrant dynamic functional connectome for depression, multimodal covarying patterns for brain function and structure, clinical diagnosis, and the clinical subtypes for depression using fMRI. The main contributions are summarized as below:

1. First, based on group information guided independent component analysis and a sliding window method, we explored disrupted topological organization and graph properties of dynamic functional network connectivity (dFNC) in major depressive disorder (MDD). Five dynamic functional states were identified, and MDD patients spent much more time in a weakly-connected State 2, which includes regions previously associated with self-focused thinking, a representative feature of depression. Additionally, MDD patients exhibited significantly reduced global efficiency, harmonic centrality, and node strength in the weekly-connected state. The aberrant FNCs in MDD were mainly related to the prefrontal, sensorimotor, and cerebellum networks, especially three commonly identified dFNCs in different states which were correlated with depressive symptom severity and cognitive performance. This study is the first attempt to investigate the dynamic functional abnormalities in MDD in a Chinese population with a relatively large sample size, which provides new evidence on aberrant time-varying brain activity and its network disruptions in MDD, which might underscore the impaired cognitive functions in this mental disorder.

2. Based on multi-site canonical correlation analysis, dynamic functional connectivity and grey matter volume for depression were jointly analyzed to investigate the multimodal covarying patterns in three independent datasets. One significantly covarying patterns was identified in three independent datasets respectively, which was related with aberrant frontoparietal connectivity, reduced grey matter volume in dorsal lateral prefrontal cortex, hippocampus, putamen, thalamus, and inferior temporal gyrus. Compared to static functional connectivity, reduced grey matter volume in hippocampus, and thalamus were significantly associated with dynamic functional connectivity. This study revealed a potential function-structure covarying patterns via data mining, which may provide new insight into the understanding of altered functional-structural coupling of large-scale brain networks in depression.

3. Motivated by the convolutional neural networks and prototype learning, we developed a brain-network-based convolutional prototype learning (BNCPL) model, which can learn representations that simultaneously maximize inter-class separation while minimize within-class distance. Saliency map was also incorporated to evaluate the contribution of input features. Functional connectivity (FC) extracted from resting-state fMRI data was used to distinguish depressive patients from healthy controls, achieving classification accuracies of 71.0% in multi-site pooling classification, with 2.4-7.2% increase compared to 3 traditional classifiers and 2 deep neural networks. Saliency map suggested that the most discriminative FCs learned by the model was primarily located in subcortical, frontoparietal, and default mode networks, especially the prefrontal-subcortical circuit, which were correlated with disease severity and cognitive ability as well. In summary, we proposed a novel BNCPL model, which improved the model interpretability and performance, suggesting that the dysregulation of the prefrontal-subcortical circuit may play a pivotal role in MDD classification, which may shed new light on understanding the pathophysiology of depression.

4. Based on a large multisite dataset (n=971), we used a data-driven, dual-clustering method called non-negative matrix factorization to identify subtypes of depression, and further investigated the diagnosis of different subtypes. The treatment-responsive rates of dfferent subtypes were explored on two independent electroconvulsive therapy (ECT) treatment datasets. Three within-dataset reproducible and cross-dataset generalizable subtypes of depression were determined, which can be defined as the fronto-parietal, fronto-occipital, and contra-lateral network subtypes according to their FC distribution. Interestingly, the fronto-parietal subtype was associated with the anxiety symptoms measured by the Hamilton anxiety scale. The subtype classification achieved an average high accuracy of 89.2% for the across-dataset prediction. Finally and most importantly, in both ECT datasets, the partial ECT response rate of the fronto-occipital subtype were 100%, much higher than the other two subtypes (30%~70%). This is the first attempt to identify and replicate MDD subtypes using functional connectivity patterns across diverse population, with rigorous cross-cohort validation and ECT treatment outcome comparison. This work revealed that different domain connectomes in depression subtypes were implicated in different clinical characteristics, which may greatly benefit the personalized intervention and help improve the prognosis.