Hepatocellular carcinomas primarily develop in livers affected by chronic liver disease. In non-alcoholic fatty liver disease (NAFLD), which affects approximately a third of the global population, hepatocarcinogenesis can occur even in early stages of disease not covered by screening practices. Novel biomarkers that evaluate those pathological aspects involved with both fatty liver disease progression and hepatocarcinogenesis — oxidative stress and associated vascular change — are needed to identify those patients most at risk. Doppler imaging enables assessment of small intrahepatic and intratumoural vascular structures, the architecture of which changes due to oxidative stress.

A database was created based on data obtained during a prospective clinical trial, containing contrast enhanced- and Doppler based ultrasound videos of the hepatic parenchyma and intrahepatic tumours. Cases and images were labelled in detail with various clinical biomarkers for disease severity and expert interpretation of images, thus facilitating the selection of relevant data and statistical analysis thereof. The concentration of various circulating proteins that are associated with neoangiogenesis were analysed in cases of varying disease severity and were developed into a label for angiogenic activity through unsupervised clustering methods.

Motion correction was applied, and feature extraction methods were developed to enable numerical description of vascular morphology and dynamic changes. Supervised learning methods were then used to evaluate the utility of these features in identifying disease, differentiating severity of disease, and identifying cancer patients based on parenchymal vascular structures.

I conclude that it is possible to obtain high quality images of sub-millimetre vascular structures within the hepatic parenchyma and intrahepatic tumours using ultrasound-based methods, and to do so consistently and reliably with clinically available equipment and software. Furthermore, descriptors of vascular morphology and flow, derived from such imaging, are able to differentiate between different stages of disease severity and between cancer and surrounding parenchyma, indicating the potential clinical relevance of these images.

Optimisation of both image collection and feature extraction methods, or application of ensemble- or deep learning methods could result in a viable tool to measure liver disease severity in certain clinical contexts. Crucially, while these vascular features correlate with assessments of fibrosis, they measure an entirely different aspect of pathology — the arterialisation and aberrant vessel formation associated with oxidative stress. Correlation with histology, and appropriate biomarkers of oxidative stress is required to further understand the relation between visualised vascular changes and pathology, and longitudinal follow-up is required to understand if vascular changes could predict hepatocarcinogenesis in NAFLD.