Methods for brain MRI image synthesis and analysis

QC 2024-12-02

Abstract

Magnetic resonance imaging (MRI) technology has made it possible to observe the inside of the brain in vivo, providing a window for researchers to better understand the processes of human brain development and aging, as well as neurological diseases and their underlying mechanisms. With the development of human brain imaging technology and the accumulation of neuroimaging data, the demand for automated computational algorithms has also increased. The success of machine learning technology in the field of computer vision has promoted its application and development in brain imaging. However, the field of medical imaging faces a series of unique challenges that have not been systematically addressed by the broad deep learning research community. In this thesis, the challenges faced in the context of three specific neuroimaging research problems are tackled. For each problem, the specific challenges are first explained, followed by the proposed solutions.

First, I focus on the specific challenges of using neuroimaging for research in the field of Alzheimer's disease (AD). In the study of AD, one difficulty is to distinguish the effects of normal aging and disease on the anatomical structure of the brain to achieve more accurate and earlier identification. Longitudinal data are an important source for understanding how normal aging affects the brain. However, the cost of collecting longitudinal data can be prohibitive, and current public longitudinal datasets still suffer from inconsistent time intervals and missing scans due to logistical constraints. To address this, I propose in the first study a diffeomorphic registration-based approach to fill in missing time points within current longitudinal datasets. To help distinguish between normal aging and disease-induced morphological changes, I propose in the second study a generative framework that can predict future brain MRI changes along healthy or pathological trajectories based on any given MRI scan. This framework is designed with diffeomorphic constraints to ensure the topological consistency of predicted morphological changes. Finally, in the third study I develop and validate a new deformation-based morphometry (DBM) framework that leverages registration-derived deformation fields for comparisons between healthy and AD groups. This new framework explicitly decomposes deformation into components attributable to normal aging and disease, providing a more nuanced understanding of their interaction.

Secondly, I focus on the pressing challenges in pediatric brain tumor segmentation. Due to the rarity of pediatric brain tumors and the high level of expertise required for tumor labeling—especially when multi-modality imaging is necessary—creating accurate automated segmentation algorithms is challenging. In response, I propose in the fourth study an unsupervised domain adaptation (UDA) framework for segmentation, utilizing relatively abundant adult brain tumor data and labels to assist in pediatric tumor segmentation.

Finally, in this thesis, I tackle the challenge of accurately registering longitudinal neuroimaging. Registration is the most fundamental and core task in the neuroimaging analysis pipelines, and its accuracy often has a serious impact on an array of subsequent analysis steps and downstream tasks. In longitudinal neuroimaging analysis, rigid registration is commonly employed to align scans obtained from the same individual. However, the deep learning-based registration community has largely focused on cross-sectional registration applications, with insufficient accuracy observed in longitudinal data. For this reason, I optimize a state-of-the-art model in the fifth study to make it more accurate for longitudinal registration.

urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-357014