Brudfors, Mikael;
(2020)
Generative Models for Preprocessing of Hospital Brain Scans.
Doctoral thesis (Ph.D), UCL (University College London).
Preview |
Text
thesis.pdf - Submitted Version Download (38MB) | Preview |
Abstract
I will in this thesis present novel computational methods for processing routine clinical brain scans. Such scans were originally acquired for qualitative assessment by trained radiologists, and present a number of difficulties for computational models, such as those within common neuroimaging analysis software. The overarching objective of this work is to enable efficient and fully automated analysis of large neuroimaging datasets, of the type currently present in many hospitals worldwide. The methods presented are based on probabilistic, generative models of the observed imaging data, and therefore rely on informative priors and realistic forward models. The first part of the thesis will present a model for image quality improvement, whose key component is a novel prior for multimodal datasets. I will demonstrate its effectiveness for super-resolving thick-sliced clinical MR scans and for denoising CT images and MR-based, multi-parametric mapping acquisitions. I will then show how the same prior can be used for within-subject, intermodal image registration, for more robustly registering large numbers of clinical scans. The second part of the thesis focusses on improved, automatic segmentation and spatial normalisation of routine clinical brain scans. I propose two extensions to a widely used segmentation technique. First, a method for this model to handle missing data, which allows me to predict entirely missing modalities from one, or a few, MR contrasts. Second, a principled way of combining the strengths of probabilistic, generative models with the unprecedented discriminative capability of deep learning. By introducing a convolutional neural network as a Markov random field prior, I can model nonlinear class interactions and learn these using backpropagation. I show that this model is robust to sequence and scanner variability. Finally, I show examples of fitting a population-level, generative model to various neuroimaging data, which can model, e.g., CT scans with haemorrhagic lesions.
Archive Staff Only
 |
View Item |
