Quantifying atherosclerosis in vasculature using ultrasound imaging

Cerebrovascular disease accounts for approximately 30% of the global burden associated with cardiovascular diseases [1]. According to the World Stroke Organisation, there are approximately 13.7 million new stroke cases annually, and just under six million people will die from stroke each year [2]. The underlying cause of this disease is atherosclerosis – a vascular pathology which is characterised by thickening and hardening of blood vessel walls. When fatty substances such as cholesterol accumulate on the inner linings of an artery, they cause a progressive narrowing of the lumen referred to as a stenosis. Localisation and grading of the severity of a stenosis, is important for practitioners to assess the risk of rupture which leads to stroke. Ultrasound imaging is popular for this purpose. It is low cost, non-invasive, and permits a quick assessment of vessel geometry and stenosis by measuring the intima media thickness. Research is showing that 3D monitoring of plaque progression may provide a better indication of sites which are at risk of rupture. Various metrics have been proposed. From these, the quantification of plaques by measuring vessel wall volume (VWV) using the segmented media-adventitia boundaries (MAB) and lumen-intima boundaries (LIB) has been shown to be sensitive to temporal changes in carotid plaque burden. Thus, methods to segment these boundaries are required to help generate VWV measurements with high accuracy, less user interaction and increased robustness to variability in di↵erent user acquisition protocols.ii This work proposes three novel methods to address these requirements, to ultimately produce a highly accurate, fully automated segmentation algorithm which works on intensity-invariant data. The first method proposed was that of generating a novel, intensity-invariant representation of ultrasound data by creating phase-congruency maps from raw unprocessed radio-frequency ultrasound information. Experiments carried out showed that this representation retained the necessary anatomical structural information to facilitate segmentation, while concurrently being invariant to changes in amplitude from the user. The second method proposed was the novel application of Deep Convolutional Networks (DCN) to carotid ultrasound images to achieve fully automatic delineation of the MAB boundaries, in addition to the use of a novel fusion of amplitude and phase congruency data as an image source. Experiments carried out showed that the DCN produces highly accurate and automated results, and that the fusion of amplitude and phase yield superior results to either one alone. The third method proposed was a new geometrically constrained objective function for the network's Stochastic Gradient Descent optimisation, thus tuning it to the segmentation problem at hand, while also developing the network further to concurrently delineate both the MAB and LIB to produce vessel wall contours. Experiments carried out here also show that the novel geometric constraints improve the segmentation results on both MAB and LIB contours. In conclusion, the presented work provides significant novel contributions to field of Carotid Ultrasound segmentation, and with future work, this could lead to implementations which facilitate plaque progression analysis for the end�user