| Summary: | Colorectal cancer (CRC) is the second largest cause of cancer deaths in the UK, with approximately 16,000 per year. Over 41,000 people are diagnosed annually, and 43% of those will die within ten years of diagnosis. The treatment of CRC patients relies on pathological examination of the disease to identify visual features that predict growth and spread, and response to chemoradiotherapy. These prognostic features are identified manually, and are subject to inter and intra-scorer variability. This variability stems from the subjectivity in interpreting large images which can have very varied appearances, as well as the time consuming and laborious methodology of visually inspecting cancer cells. The work in this thesis presents a systematic approach to developing a solution to address this problem for one such prognostic indicator, the Tumour:Stroma Ratio (TSR). The steps taken are presented sequentially through the chapters, in order of the work carried out. These specifically involve the acquisition and assessment of a dataset of 2.4 million expert-classified images of CRC, and multiple iterations of algorithm development, to automate the process of generating TSRs for patient cases. The algorithm improvements are made using conclusions from observer studies, conducted on a psychophysics experiment platform developed as part of this work, and further work is undertaken to identify issues of image quality that affect automated solutions. The developed algorithm is then applied to a clinical trial dataset with survival data, meaning that the algorithm is validated against two separate pathologist-scored, clinical trial datasets, as well as being able to test its suitability for generating independent prognostic markers.
|
|---|
