- 期刊
使用LSTM深度神經網路與靜息態功能性磁振造影探索可逆性腦血管收縮症候群
Exploring Reversible Cerebral Vasoconstriction Syndrome using LSTM Deep Neural Networks and Resting-state Functional Magnetic Resonance Imaging
摘要
可逆性腦血管收縮症候群（Reversible Cerebral Vasoconstriction Syndrome, RCVS）是一種可逆的腦動脈節段性和多發性血管收縮。我們推測腦血管張力自主調節失衡可能是此病生理機轉的主軸，血管張力的突然改變可能造成節段性腦血管收縮及遠端小動脈代償性的血管擴張，後者可能涉及到血管周圍的疼痛神經，導致突發性頭痛。而與血管自主神經網絡調節失衡相關的因素可能包括交感神經活化過度增加、氧化壓力、以及內皮細胞功能異常等多重因素的共同作用。靜息態功能磁共振成像（Resting-state functional Magnetic Resonance Imaging, rs-fMRI）是目前常用的臨床磁振影像技術之一，其具有非侵入性、可重複使用的優勢，且無需病患執行複雜的指令。在大腦靜息狀態下，出現低頻血氧水平依賴性（Blood Oxygenation Labeled Dependent, BOLD）信號，這種信號與神經元活化之間的關聯性非常高。大腦在靜息狀態下的活動被視為一種連續自主的腦活動，因此，rs-fMRI數據代表了大腦區域的靜息動態活動情況。假設這些時間序列訊號中神經元活化的信號波動，在不同大腦功能區域的活動狀態也會因為疾病發生變化，這比其他使用靜息態功能磁振造影的功能連結分析方法更直接與直觀。為了處理這些時間動態數據，我們採用深度學習中長短期記憶網絡（Long Short Term Memory networks, LSTMs），這是一種特殊類型的時間循環神經網絡（Recurrent Neural Network, RNN）。由於其獨特的結構，LSTM非常適合處理和預測時間序列中的長間隔和延遲事件訊號。我們使用LSTM深度神經網絡處理rs-fMRI時間序列數據，根據自律神經和其他功能網絡等相對應的腦區，直接擷取這些腦區的rs-fMRI時間序列訊號作為學習和訓練的數據，以建立模型，進而對RCVS患者和健康對照者（Healthy Controls, HC）進行分類。
並列摘要
Reversible Cerebral Vasoconstriction Syndrome (RCVS) is a reversible condition characterized by segmental and multifocal constriction of cerebral arteries. We hypothesize that an imbalance in the autonomic regulation of vascular tone may be central to the pathophysiology of this condition, with abrupt changes in vascular tone leading to segmental cerebral vasoconstriction and compensatory vasodilation of distal arterioles, possibly involving perivascular pain fibers, and resulting in sudden-onset headaches. Dysregulation of the autonomic nervous system, possibly driven by factors such as excessive sympathetic nerve activity, oxidative stress, and endothelial dysfunction, may play a role in this process. Resting-state functional Magnetic Resonance Imaging (rs-fMRI) is a commonly used clinical magnetic resonance imaging sequence that offers non-invasiveness, repeatability, and eliminates the need for complex patient instructions. During the resting state of the brain, there is a high correlation between low-frequency Blood Oxygenation Labeled Dependent (BOLD) signals and neuronal excitability. The brain's activity during resting state is considered a continuous, spontaneous brain function. Therefore, rs-fMRI data represents dynamic brain activity, and we hypothesize that the activity states of different brain regions change over time based on the signal fluctuations of neuronal activation in the time series, providing richer information than static functional connectivity analysis alone. To process this data, we employ Long Short Term Memory networks (LSTMs) of deep learning, a type of Recurrent Neural Network (RNN) known for its ability to handle and predict long intervals and delays within time series data. In this study, we utilize LSTM deep neural networks to process rs-fMRI time series data. We extract data directly from specific brain regions corresponding to the autonomic nervous system and other functional networks. These time series data serve as inputs for our LSTM model, allowing us to build a classification model to distinguish between RCVS patients and Healthy Controls (HC).