脑深部电刺激（Deep Brain Stimulation, DBS）是治疗帕金森病（Parkinson's Disease, PD）的有效手段，但其调控作用的脑功能机制仍不清楚。这阻碍了对DBS疗法的优化及适应症的拓展。功能磁共振成像（functional Magnetic Resonance Imaging，fMRI）因具有全脑高空间分辨率功能信号同步采集的优点，被认为是研究DBS调控机制的关键工具。得益于近年来磁共振兼容电极的研发成功，高场强下的DBS fMRI采集成为了现实。但DBS的植入会给fMRI数据引入噪声和信号缺失等问题，限制了个体fMRI数据的可靠性和可用性。建立PD患者的脑功能连接与运动症状的关联是研究DBS调控机制的重要基础。症状和脑功能的个体差异使得传统的组平均分析方法难以准确地捕捉到这种关联。因此，精细化解析脑功能个体差异并开发出个体化fMRI的分析方法是建立脑功能与症状关联的前提。为此，本文主要做了以下工作： 首先，本文开发了基于脑外成分分解的静息态fMRI降噪算法和基于深度学习的fMRI缺失信号重建算法。降噪算法降低了DBS植入导致的脑内外噪声，提升了个体内的fMRI数据的可靠性，改善了被试间的功能连接组的相似性。重建算法重建了DBS植入后导致的个体fMRI缺失信号，提高了数据的可用性。 其次，在对信号的降噪和重建的基础上，本文以听觉皮层为例研究了局部脑区的个体间脑功能差异和细粒度功能区剖分。利用静息态功能连接揭示了人和猕猴听觉皮层的个体间功能差异，发现了听觉皮层存在高低个体差异的两个区域，并且个体差异的特征在物种间存在一致性。利用听觉皮层的个体差异分布，进一步地对听觉皮层进行了细粒度功能区剖分。 再次，在局部脑区功能区剖分的启发下，本文以“分而治之”的思路实现了全脑细粒度功能区剖分，并开发了相应的个体化算法。该算法在个体内可重复性高并揭示了个体差异，可以准确捕捉任务激活的个体差异。基于该方法，还发现了前额叶精细的脑功能梯度，并发现该梯度在纹状体上存在对应关系。 最后，本文初步探索了利用个体化细粒度功能区剖分得到的个体功能连接，建立了可泛化的机器学习模型，该模型可以稳健预测出PD患者的运动症状严重程度，并可以泛化到模型未见过的DBS数据集中。利用该方法找到的影像学标记物可以揭示出DBS高频刺激引起的功能连接正常化效果。
Deep Brain Stimulation (DBS) is an effective neuromodulatory treatment for Parkinson's Disease (PD) and a potential intervention for a variety of brain disorders. However, the neural mechanisms of effectiveness remain unclear. The lack of clarity on the mechanism is a serious obstacle to the optimization of DBS therapies and the expansion of indications. Functional magnetic resonance imaging (fMRI) is considered to be an appropriate tool to unravel the regulatory mechanisms of DBS due to its capability to sample functional signals at high spatial resolution throughout the whole brain. Thanks to the successful development of high-field MRI-compatible electrodes in recent years, prolonged fMRI acquisition in the ON state of DBS has become a reality. However, fMRI data from patients with implanted DBS electrodes have problems such as high noise and fMRI signal loss, which limit the reliability of intra-individual fMRI data and the analysis of critical functional areas. Due to the existence of individual differences in brain functional architecture, the traditional group-average analysis method blur the fine-grained brain functional organization and thus missing subtle but critical findings. Therefore, developing individualized fMRI analytic methods is the basis for revealing the regulatory mechanisms of DBS. The following studies were conducted in the thesis. First, the thesis developed a resting-state fMRI (rsfMRI) noise correction algorithm based on noise component decomposition and a compromised signal reconstruction algorithm based on deep learning. The noise reduction algorithm reduced the whole brain noise induced by DBS implantation, improved the reliability of intra-individual fMRI data, and increased the similarity of functional connectome between subjects. fMRI compromised signal reconstruction algorithm reconstructed the loss signal induced by DBS implantation for each patient, and improved the usability of data. Second, based on noise reduction and signal reconstruction, this thesis investigated the inter-individual functional differences in local cortical areas and fine-grained functional parcellation was performed. Results revealed inter-individual functional differences in human and macaque auditory cortices using resting-state functional connectivity, and identifies two regions of auditory cortex with high and low individual variability. The characteristics of individual differences were consistent across species. The distribution of individual variability in the auditory cortex was used to further parcellated the auditory cortex into fine-grained functional regions. In addition, inspired by the fine-grained functional area parcellation of the auditory cortex, we extended the method to the whole cerebral cortex. Additionally, an individual-specific functional parcellation algorithm was developed to achieve high intra-individual reliability and revealed inter-individual variability. Individual parcellation based on rsfMRI can accurately capture inter-individual differences in task-evoked activations. Importantly, we found the fine-grained functional gradient organization in the prefrontal cortex and the correspondence between the prefrontal cortex and the striatum. Finally, we initially explored individual-specific functional connectivity estimated from individualized fine-grained functional parcellation to build a generalizable machine learning model. The model could robustly predict the severity of motor symptoms in PD patients. The imaging markers revealed normalization in functional connectivity due to high-frequency DBS.
