利用磁振造影、圖論分析和圖深度學習解譯大腦網路：在認知和臨床神經科學中的應用

Deciphering Complex Brain Networks with Magnetic Resonance Imaging, Graph-theoretical Analysis and Deep Learning on Graphs: Applications in Cognitive and Clinical Neuroscience

林奭彥(Shih-Yen Lin)

指導教授 ： 陳永昇 ； 郭立威

國立交通大學/資訊學院/資訊科學與工程研究所/博士(2020年)

摘要

我們的認知與大腦區域之間的聯繫和互動緊密相關，而磁振造影（magnetic resonance imaging, MRI）的發展使得人們能夠以非侵入的方式觀察大腦的連接，引起了諸多關於大腦連接性的研究。近年來，出現了許多新穎的大腦網路分析方法。圖論分析利用網路科學的理論，並使用各種網路指標來描述高階網路屬性。圖信號處理旨在將一般信號處理的理論推廣到不規則的數據域。受到圖形信號處理概念的啟發，圖深度學習是一種推廣到不規則數據域的深度學習模型，可以更好地說明區域間的交互作用。但這些新穎的分析方法可能尚未廣泛應用於腦網路分析，並且仍然存在各種未解決的問題。因此，本論文旨在解決這些局限性，並探討上述分析方法在基於模型和數據驅動的腦科學研究中的潛力。
本論文包含三項研究。在第一個研究中，我們將多種結構和功能連接性指標整合到了圖論分析框架中，並研究了輕度認知障礙（mild cognitive impairment, MCI）和阿茲海默症　（Alzheimer’s disease, AD）患者腦網路結構的變化。與一般方法相比，這種用於大腦網路分析的多參數方法反映了有關MCI或AD中大腦網路結構變化的更多信息。在我們的研究中，除了局部效率的廣泛變化外，還發現MCI和AD中的結構網路結構發生了重大破壞，主要發生在邊緣系統，額葉和枕葉區域。在全腦範圍下，我們的結果顯示，結構網路的破壞在從MCI到AD階段的不同連接性指標和全域網路指標之間都是一致的。我們在邊緣系統中，還發現連接性及擴散速率指標發生了顯著變化。我們的發現表明，特定於管道的指標，例如各向異性(anisotropy)和擴散速率（diffusivity），比基於纖維數量(streamline count)的指標提供了更敏感和可解釋的指標。此外，使用反向徑向擴散率(inversed radial diffusivity) 還提供了更多信息，以了解由AD進程引起的網路結構變化及其可能的原因。我們得出的結論是，此研究提議的多參數網路分析可能有助於早期MCI診斷和AD預防。
在我們的第二項研究中，我們通過使用圖論網路分析法研究了語音情緒不同屬性下的功能性大腦網路的變化。我們從功能性磁振造影數據計算了不同程度的情緒喚起度(arousal)和正負向性(valence)下的語音情緒刺激下以及靜息狀態下的圖論網路指標。我們的統計結果表明，大腦網路在全腦、區域和連接性級別上表現出網路屬性的顯著變化，尤其是在具有高喚起度或負向的語音情緒刺激下。通過這項研究，我們對理解語音情緒如何調節大腦網路有了更多的了解。這些發現可能會揭示人腦如何處理情感言語以及它如何區分不同的情感狀況。
在第三項研究中，受關於動態功能連接性的神經科學研究和幾何深度學習的最新發展的啟發，我們提出了一種腦信號的新型的深層神經網路分類框架，該框架可以學習動態功能連通性之表示(representation)。首先，我們提出了新穎的自適應編碼器(adaptive edge encoder)，該編碼器可學習動態功能連通性的表示形式，以優化分類性能。其次，以生物學啟發的分層方式執行腦信號的表示學習，與傳統的卷積神經網路(convolutional neural network, CNN) 模型相比，可以進行更直觀的解釋。在使用人腦連接體計畫(Human Connectome Project)的靜息狀態功能性磁振造影資料的性別分類實驗中，我們所提出的模型在分類準確性方面優於文獻上最先進的模型。此外，使用梯度加權類激活映射(gradient class activation mapping) 的可視化還顯示了與性別差異相關的大腦區域，並與神經學研究的發現一致。

關鍵字

大腦網路分析 ； 磁振造影 ； 圖論分析 ； 圖深度學習

並列摘要

Our cognitions are closely associated with the connections and interactions between brain regions. In recent decades, the development of magnetic resonance imaging (MRI) has enabled in vivo observations of the brain connections, and has given rise to many researches on brain connectivity. Many novel methods for brain network analysis have also emerged in recent years. Graph-theoretical analysis utilizes the knowledge in network science and describes high-level network attributes using various network measures. Graph signal processing aims to generalize conventional signal processing operations to the irregular graph domain. Inspired by the concept of graph signal processing, deep learning on graphs are deep learning models generalized to the irregular data domain and can better account for interregional interactions. However, these novel analysis methods has yet been widely applied to brain network analysis, and various open issues still exist. This thesis therefore serves as our attempts to address these limitations, and explore the potentials of the aforementioned analysis methods in the study of model-based and data-driven brain science.
This thesis contains three studies. In the first study, we incorporated multiple structural and functional connectivity metrics into a graph theoretical analysis framework and investigated alterations in brain network topology in patients with mild cognitive impairment (MCI) and Alzheimer’s Disease (AD). This multiparametric approach for brain network analysis may reflect additional or complementary information regarding the topological changes in brain networks in MCI or AD, when compared with conventional methods. In our study, significant disruption of structural network topology in MCI and AD was found predominantly in regions within the limbic system, prefrontal and occipital regions, in addition to widespread alterations of local efficiency. At a global scale, our results showed that the disruption of the structural network was consistent across different edge definitions and global network metrics from the MCI to AD stages. Significant changes in connectivity and tract-specific diffusivity were also found in several limbic connections. Our findings suggest that tract-specific metrics (e.g., fractional anisotropy and diffusivity) provide more sensitive and interpretable measurements than does metrics based on streamline count. Besides, the use of inversed radial diffusivity provided additional information for understanding alterations in network topology caused by AD progression and its possible origins. We conclude that the used of this proposed multiparametric network analysis framework may facilitate early MCI diagnosis and AD prevention.
In our second study, we investigated the alterations of functional brain network under different attributes of speech emotion by using graph-theoretical network analysis. We computed high-level graph-theoretical network measures from functional MRI (fMRI) data under resting-state and vocal stimuli at different levels of arousal and valence. Our statistical results show that brain network exhibits significantly altered network attributes at global, nodal and connectivity levels, especially under vocal emotional stimuli with high arousal or negative valence. Through this study, we have gained more insights into how comprehending emotional speech modulates brain networks. These findings may shed light on how the human brain processes emotional speech and how it distinguishes different emotional conditions.
In the third study, inspired from the neuroscientific studies on dynamic functional connectivity and the latest developments in deep learning on graphs, we proposed a novel classification framework for brain signals that allowed an adaptive representation of dynamic functional connectivity. First, we proposed novel adaptive edge encoder that allow for learnable representations of dynamic functional connectivity to improve the classification performance. Second, the representation learning of the brain signal in the proposed model is performed in a biologically-inspired hierarchical manner, allowing for a more intuitive interpretation as compared to conventional convolutional neural network model. By performing a gender classification task using resting-state fMRI dataset from Human Connectome Project, we demonstrated that our model outperformed the state-of-the-art models in terms of accuracy. In addition, visualization using gradient-weighted class activation mapping also revealed several brain regions well supported by neurological studies on gender-related differences.

並列關鍵字

brain network analysis ； Magnetic Resonance Imaging ； Graph-theoretical Analysis ； Deep Learning on Graphs

參考文獻

[1] Timothy E. J. Behrens and Saad Jbabdi. Chapter 15 - MR Diffusion Tractography,

pages 333–351. Academic Press, San Diego, 2009.

[2] Karl J Friston. Functional and effective connectivity: a review. Brain connectivity,

1(1):13–36, 2011.

[3] M. Rubinov and O. Sporns. Complex network measures of brain connectivity: uses