Analysis and processing of dynamic and structural magnetic resonance imaging signals for studying small vessel disease

Abstract

Cerebral small vessel disease (CSVD) describes multiple and dynamic pathological processes disrupting the optimum functioning of perforating arterioles, capillaries and venules, increasing the risk of stroke and dementia. Although the pathogenesis of this disease is still elusive, the breakdown of the blood-brain barrier (BBB), which would hinder brain waste clearance, is thought to play a pivotal factor in it. Nonetheless, the microscopic origin and nature of these abnormalities and the lack of a ground truth make the study of CSVD in vivo in humans via magnetic resonance imaging (MRI) challenging and signal processing schemes likely to be sub-optimal. In this doctoral thesis, we proposed signal analysis and processing techniques to improve the quantification and characterisation of subtle and clinically relevant neuroimaging features of CSVD. We applied our proposals to analyses of structural and dynamic-contrast enhanced MRI (sMRI and DCE-MRI) to better characterise CSVD.

DCE-MRI is commonly used to investigate cerebrovascular dysfunction, but the extremely subtle nature of the signal in CSVD makes it unclear whether signal changes are caused by microscopic yet critical BBB abnormalities. Moreover, ethical and safety considerations in vivo and the lack of validation frameworks hinder optimising imaging protocols and processing schemes. To cope with these issues, we thus proposed an open-source computational human brain model for mimicking the four-dimensional DCE-MRI acquisition process. With it, we quantified the substantial impact of spatiotemporal considerations on permeability mapping, detected sources of errors that had been overlooked in the past, and provided evidence of the harmful effect of post-processing or lack thereof on DCE-MRI assessments. Perivascular spaces (PVS) in the brain, which are involved in brain waste clearance, can become visible in sMRI scans of patients with neuroimaging features of CSVD, but their automatic quantification is challenging due to the size of PVS, the incidence and presence of imaging artefacts, and the lack of a ground truth. We first proposed a computational model of sMRI to study and compare current PVS segmentation techniques and identify major areas of improvement. We confirmed that optimal segmentation requires tuning depending on image quality and that motion artefacts are particularly detrimental to PVS quantification. We then proposed a processing strategy that distinguished high-quality from motion-corrupted images and processed them accordingly. We demonstrated such an approximation leads to estimates that correlate better with clinical visual scores and agree more with full manual counts. After optimisation using our proposals, we also found PVS measurements were associated with BBB permeability, in accordance with the link between brain waste clearance and endothelial dysfunction.

This work provides means for understanding the effect of image acquisition and processing on the assessment of subtle markers of brain health to maximise confidence of studies of endothelial dysfunction and brain waste clearance via MRI. It also constitutes a cornerstone on which future optimisation and development can be based upon.