UCLA

Magnetic resonance imaging (MRI) is used extensively to perform a wide range of neuroscientific studies. These include research conducted on humans and animals, with mice in particular being used frequently to model human neurobiology. In these studies, MRI data are typically processed to extract measures of interest after correcting for unwanted noise, motion, and other artifacts that could lead to an inaccurate analysis of the underlying brain anatomy. These imaging workflows rely on automated tools that are often preferred over manual methods that are labor-intensive and prone to interrater variability. The accuracy and reliability of these tools are critically important to the outcomes of the analyses. However, the distribution of image processing pipelines can pose issues with respect to portability, interoperability, and reproducibility. Additionally, in the case of preclinical imaging, a limited number of tools exist that are designed specifically for analyzing mouse MRI. In this dissertation, we focused on the construction of comprehensive frameworks that ensure reproducible neuroimaging analyses with an emphasis on developing improved tools for segmentation of mouse MRI data. As a first step, we developed the BrainSuite BIDS App, which encapsulates the core MRI processing functionality of BrainSuite for analyzing human brain MRI data. The BrainSuite BIDS App follows the BIDS App standard and implements a containerized set of workflows for performing participant-level and group-level analysis on anatomical, diffusion, and functional MRI. We then adopted a similar approach to create a mouse-specific containerized analysis framework called the Mouse MRI Anatomical Pipeline BIDS App (MMAP). We achieved this by using a combination of existing tools and new tools that we developed to address important problems in mouse image analysis. We first implemented the Mouse Brain Extractor, a new method for brain extraction that builds upon the SwinUNETR architecture, to improve robustness. Our approach supplies the network model with supplementary spatial information in the form of fixed absolute positional encodings, which we call Global Positional Encodings (GPEs). We also implemented a tissue classifier by training and evaluating multiple deep-learning methods. This work culminated in the integration of these new tools and existing tools into a containerized framework for end-to-end analysis. Our system also enables switching multiple different competing approaches for different steps in the workflow, facilitating comparison of a variety of strategies for mouse MRI analysis. We applied these developments on two neuroimaging studies that analyzed gross and regional volumetric measures of the brain to examine the effects of two conditions. In a study of mice with experimental autoimmune encephalomyelitis (EAE), we observed cortical atrophy in EAE-induced mice. In another study examining the effect of Chd8 haploinsufficiency on mouse brains, we found megalencephaly in the heterozygous group. These findings corresponded generally with existing literature and demonstrate the practical utility of the methods developed in this dissertation in their application to neuroimaging studies.