| Summary: | Neuroimaging has enabled the observation of damage to the white matter that occurs frequently in elderly population and is depicted as hyperintensities in specific magnetic resonance images. Since the pathophysiology underlying the existence of these signal abnormalities and the association with clinical risk factors and outcome is still investigated, a robust and accurate quantification and characterisation of these observations is necessary. In this thesis, I developed a data-driven split and merge model selection framework that results in the joint modelling of normal appearing and outlier observations in a hierarchical Gaussian mixture model. The resulting model can then be used to segment white matter hyperintensities (WMH) in a post-processing step. The validity of the method in terms of robustness to data quality, acquisition protocol and preprocessing and its comparison to the state of the art is evaluated in both simulated and clinical settings. To further characterise the lesions, a subject-specific coordinate frame that divides the WM region according to the relative distance between the ventricular surface and the cortical sheet and to the lobar location is introduced. This coordinate frame is used for the comparison of lesion distributions in a population of twin pairs and for the prediction and standardisation of visual rating scales. Lastly the cross-sectional method is extended into a longitudinal framework, in which a Gaussian Mixture model built on an average image is used to constrain the representation of the individual time points. The method is validated through a purpose-build longitudinal lesion simulator and applied to the investigation of the relationship between APOE genetic status and lesion load progression.
|
|---|
