| Summary: | <p>Biomedical image registration is well established as a core component of most neuroimaging processing pipelines. It allows multiple subjects to be analysed using a common frame of reference, thereby facilitating both the localisation and quantification of similarities and differences between individuals or population groups.</p>
<p>An assumption inherent to such analyses is that registration produces a one-to-one mapping between common anatomical locations across subjects. Errors in image registration violate this assumption, typically reducing both the power to detect effects of interest and the ability to localise those effects. But image registration is an inherently ill-posed problem, and therefore the true, unique one-to-one mapping can generally never be found (if it even exists). Instead, registration methods can only attempt to produce the most likely deformation, given the subject data and models of what type of mappings are more probable than others.</p>
<p>In this thesis I present two approaches to improving the accuracy of nonlinear registration, particularly in the context of magnetic resonance imaging of the brain.</p>
<p>The first utilises a computationally complex regularisation model which penalises the stretching effect of a deformation in a highly nonlinear way. The method produces more anatomically plausible mappings, with less distortion of both volume and shape, when compared to current state-of-the-art methods. It achieves this while simultaneously equalling or bettering the alignment accuracy, as measured by the overlap of expert-segmented anatomical brain regions.</p>
<p>The second leverages the fact that it is increasingly common for both scalar and tensor imaging modalities, containing complementary spatial information, to be acquired for each subject in a dataset. The similarity between multiple modalities are then simultaneously compared, accounting for differences in both tissue location and orientation, to better inform the registration in a data-driven way. This leads to a measurable and significant improvement in anatomical consistency across subjects, evaluated using group-level task-based brain activation maps.</p>
<p>The cost associated with these improvements is an increase in algorithmic complexity. This is addressed by leveraging the massively parallelised computing capabilities of graphics processing units to offset the rise in complexity. To this end, I present my implementation of a new software tool, the MultiMOdal Registration Framework (MMORF), which forms the basis of all of the results herein.</p>
<p>My hope is that the benefits I present in this thesis will allow MMORF to have a real impact in unifying volumetric image registration in the field of neuroimaging.</p>
|
|---|