面向神经影像应用的可扩展且可解释的空间模型

Yifan YuKeble CollegeUniversity of Oxford A thesis submitted for the degree ofDoctor of PhilosophyHilary 2025 Acknowledgements First and foremost, I would like to express my deepest gratitude to my supervisor,Professor Thomas E. Nichols.His unwavering support, insightful guidance, andconstant inspiration have been invaluable throughout my PhD journey. It has beena true privilege to be mentored by someone I regard as a research role model. Hisdepth and breadth of knowledge have been both humbling and enlightening. A specialacknowledgement goes to Tom, whose thoughtfulness and readiness to offer supportduring challenging times, both academically and personally, have meant so much tome. I am also deeply grateful to Professor Angela Laird and Dr. Alejandro De La Vegafor their insightful feedback on my work. My heartfelt thanks go to Dr. James Kentfor guiding me in contributing to the NiMARE package, and for the countless hourshe dedicated to teaching me formalised coding practices and improving the efficiencyand readability of my codes.I would also like to express my sincere gratitude tomy internship manager, Martina De Stefani, and all my colleagues at Amazon. Thisopportunity to experience the difference between academic and industrial research,has been invaluable in helping me make decisions about my future career path. I would also like to thank Professor Ludovica Griffanti and Professor Simon Eickhofffor their insightful feedbacks and stimulating discussions during my PhD viva. Theircomments have profoundly deepened my understanding of data interpretation andwill undoubtedly shape my future research endeavours. I am also grateful for theconsiderable time they dedicated to thoroughly reviewing my thesis. My appreciationalso goes to thank Professor Saad Jbabdi and Dr Natalie Staplin for the invaluablefeedback they provided during my transfer and confirmation of status assessments.My sincere thanks also go to the wonderful labmates and colleagues of Tom’s researchgroup, past and present. In particular, I would like to express my gratitude to AnnaMenacher, Kan Keeratimahat, Lav Radosavljevic, Saba Ishrat, Angeline Lee, ThomasMaullin-Sapey, Alex Bowring, Sam Davenport, George Hustchings, Emma Prevot, Yang Sun, Anya Topiwala, Bernd Taschler, Konstantin Shestopaloff, Habib Ganjgahi,for their support, collaboration and inspiring discussions throughout my PhD journey. Moreover, I feel incredibly fortunate to be surrounded by so many amazing andcaring friends, far too many to name individually. I would like to especially thankNatalia Hong, Kevin Wang, Hang Yuan, Chenyang Wang from Keble College; XiLin, Linying Yang, Sahra Ghalebikesabi, Yixuan He, Hanwen Xing, Yanzhao Yang,Ning Miao, Chao Zhang, Yutong Lu, Zhixiao Zhu, Guneet Singh Dhillon from theDepartment of Statistics, as well as Sijia Yao, Kangning Zhang, Ziyun Liang andmany others from other departments at Oxford, I am also grateful for friends fromearlier chapters of my life, including those from high school or undergraduate studies –Zhiqi Wang, Kaiyue Zhang, Tianxiao Wang, Yongtong Chen, Wenxuan Dong, LuyangCui, Yuru Bai, Zhimeng Shi and Xue Lin. Finally, above all, my deepest thanks go to Jin Xu for his love and companionship,for the countless train tickets between Cambridge and Oxford, between Tübingenand Luxembourg. I couldn’t imagine my PhD journey without him. He has broughtimmense joy, unwavering support and continual inspirations to both my life andresearch. Additionally, I am eternally grateful to my parents Zhisan Yu and Xiao Xia.Their unconditional support and constant encouragement gave me the freedom topursue my passions, and the thought of them has always brought me comfort duringmy most difficult times. Abstract Neuroimaging techniques, such as functional magnetic resonance imaging (fMRI) andstructural MRI, have become essential tools for understanding function and pathologyof the human brain.fMRI allows researchers to identify brain regions associatedwith specific cognitive and behavioural processes by measuring blood-oxygen-level-dependent (BOLD) signals.Structural MRI provides high-resolution mapping ofbrain anatomy, allowing for the identification of morphological alternations, such aswhite matter lesions, which improves our understanding of neurodegenerative andcerebrovascular diseases. Despite recent developments in data sharing and availabilityof large-scale neuroimaging cohorts, several common analytical challenges remain.Individual neuroimaging studies often rely on relatively small sample sizes, whichlimits statistical power and reduces the generalisability of findings. Many existinganalytical approaches struggle with balancing model complexity, interpretability andscalability. Another critical limitation is the lack of methods that explicitly modelspatial dependence in neuroimaging data while maintaining computationally efficiency.Addressing these challenges is crucial for improving the reliability, reproducibility, andclini