Quantitative histopathological assessment of liver fibrosis

Thesis

Metabolic-associated fatty liver disease (MAFLD) affects an estimated 25% of the world’s population. The feature of MAFLD most strongly associated with adverse outcomes is liver fibrosis. Severe fibrosis is associated with an increased risk of adverse cardiovascular events, the development of malignancy and end-stage liver disease. Pathology staging systems for assessment of the structural changes resulting from liver fibrosis suffer from poor reproducibility. Digital pathology offers ways of quantifying fibrotic changes through simple imaging markers such as collagen proportionate area (CPA). The predictive accuracy of such markers is largely dependent on the quality of the tissue pre-examination processes (i.e histological staining). In this work, we propose a new way of estimating measurement uncertainty of collagen proportionate area through the employment of deep-learning ensembles in a Bayesian inference framework. Nest, we design a novel graph-neural network pipeline for automatic staging of liver fibrosis into an established pathological system. Finally, we conduct morphological subtyping of fibrosis patterns in an unsupervised way to develop a novel representation of liver fibrosis at the level of local cellular neighbourhoods. We show how this representation can be integrated with the current imaging markers, such as CPA, to form a succinct visual summary of the current disease state. Finally, we explore the possibility of applying the developed representation to tracking the progression and regression of disease over time.

Authors: Wojciechowska, M

• DOI: 10.5287/ora-xoqmwj4dd
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: liver histology; non-alcoholic fatty liver disease; digital pathology; metabolic-associated fatty liver disease; deep learning; quantitative pathology; liver fibrosis
• Subjects: Liver; Fibrosis; Biomedical engineering; Deep learning (Machine learning)