The application of artificial intelligence and image analysis to novel prognostic classification systems of colorectal cancer

Colorectal cancer (CRC) is the third most common form of cancer in the world and the second most common cancer related mortality. Adequate staging of CRC is important for understanding patient prognosis and determining appropriate therapy regimens. CRC staging is currently performed according to the Tumour, Node, Metastasis (TNM) staging criteria, which has remained the gold standard around the world since its introduction. However, the variable prognosis of Stage II/node negative disease and uncertainty around best therapeutic practices for these patients has been a continuing issue with the TNM system, one that is still yet to be adequately addressed. Through extensive research, novel prognostic features, assessed on diagnostic Haematoxylin & Eosin (H&E) sections and through simple Immunohistochemistry (IHC), have been shown to supplement the TNM criteria and address this unmet clinical need. Furthermore, novel classification systems that incorporate multiple features of the Tumour Microenvironment (TME) and assign patients to independent groups have been developed and often stratify patients for prognosis better than the TNM system, as well as providing additional prognostic and theragnostic information. The adoption of these novel factors and classification systems into clinical pathology has been hindered by persistent interobserver variability and a lack of clear and standardised assessment criteria. Image analysis presents a means to reduce the subjectivity of these criteria and increase their potential clinical utility. With the advent of artificial intelligence and its continued development within image analysis, the automated assessment of novel features on clinical sections is becoming increasingly reliable and reproducible. Therefore, this thesis aims to utilise image analysis and artificial intelligence to automate the assessment of specific features of the CRC TME, assess the prognostic utility of novel TME features individually and in combination, and compare the performance of digital assessment to human assessment. The Glasgow Microenvironment Score (GMS) is a combined assessment of the stromal density of the tumour, quantified through the Tumour Stroma Percentage (TSP), and the peritumoural inflammatory response, assessed using the Klintrup – Mäkinen (KM) criteria, that assigns patients to one of three individually prognostic groups. Using a Convolutional Neural Network (CNN) to semantically segment H&E Whole Slide Images (WSI) and quantify the tumour associated stroma demonstrated that image analysis is able to reliably conduct TSP assessment across multiple retrospective patient cohorts and a large clinical trial cohort, prognostically stratify these patients according to the TSP criteria, and outperform human assessment for prognostic significance. Image analysis quantification of peritumoural lymphocyte density on H&E WSI with manually annotated invasive margins showed a significant association with prognosis, comparable to that seen in manual KM assessment, again across multiple cohorts. Combining both image analysis approaches according to the GMS criteria outperformed pathologist assessment for survival stratification, highlighting the ability of image analysis algorithms to reliably perform individual assessments and retain the prognostic significance when used in combination. The Phenotypic Subtypes of CRC are a translation of the phenotypic signatures of the Consensus Molecular Subtypes (CMS) to tissue – based assessment, incorporating Ki67 IHC into the GMS criteria with the TSP and KM. Ki67 expression is utilised to further stratify the GMS group with intermediate prognosis, providing additional information about the TME. To quantify Ki67 expression, a CNN was again used to semantically segment Tissue MicroArray (TMA) cores stained for Ki67 via IHC and the percentage of Ki67+ tumour cells was determined using an automated, CNN – based cell detection algorithm. Ki67 expression determined through automated analysis was significantly associated with prognosis individually, and when combined with the TSP and KM criteria, the Phenotypic Subtypes determined through image analysis were highly prognostic again across multiple cohorts. Furthermore, the image analysis subtypes identified a group of patients with a chemotherapy dependent improvement in survival, demonstrating the clinical utility of image analysis for determining patient prognosis and potentially guiding therapy regimens. The data presented in the current thesis demonstrates that image analysis is able to reliably and reproducibly assess novel features of the TME from clinical WSI, perform these assessments across multiple independent patient cohorts, significantly stratify patients for prognosis, and has the potential to be utilised in clinical pathology to aid therapeutic decisions and improve patient outcomes.