Statistical analysis for longitudinal MR imaging of dementia
Serial Magnetic Resonance (MR) Imaging can reveal structural atrophy in the brains of subjects with neurodegenerative diseases such as Alzheimer’s Disease (AD). Methods of computational neuroanatomy allow the detection of statistically significant patterns of brain change over time and/or over multiple subjects. The focus of this thesis is the development and application of statistical and supporting methodology for the analysis of three-dimensional brain imaging data. There is a particular emphasis on longitudinal data, though much of the statistical methodology is more general. New methods of voxel-based morphometry (VBM) are developed for serial MR data, employing combinations of tissue segmentation and longitudinal non-rigid registration. The methods are evaluated using novel quantitative metrics based on simulated data. Contributions to general aspects of VBM are also made, and include a publication concerning guidelines for reporting VBM studies, and another examining an issue in the selection of which voxels to include in the statistical analysis mask for VBM of atrophic conditions. Research is carried out into the statistical theory of permutation testing for application to multivariate general linear models, and is then used to build software for the analysis of multivariate deformation- and tensor-based morphometry data, efficiently correcting for the multiple comparison problem inherent in voxel-wise analysis of images. Monte Carlo simulation studies extend results available in the literature regarding the different strategies available for permutation testing in the presence of confounds. Theoretical aspects of longitudinal deformation- and tensor-based morphometry are explored, such as the options for combining within- and between-subject deformation fields. Practical investigation of several different methods and variants is performed for a longitudinal AD study.