Methods for tumour aggression prediction in colorectal cancer through virtual immunohistochemistry, supervised, and self-supervised deep learning

Walsh, Christopher David (2024) Methods for tumour aggression prediction in colorectal cancer through virtual immunohistochemistry, supervised, and self-supervised deep learning. PhD thesis, University of Glasgow.

Abstract

Tumour budding has emerged as a critical, independent predictor of survival and regression in colorectal cancer. It has elucidated the aggressive nature of certain tumours and the survival heterogeneity among tumours of a similar stage as determined by the TNM system. However, its clinical integration has been hampered by challenges like the absence of a standardised scoring methodology and inter-observer variability when scoring in H&E. This work presents three methods to assist or automate tumour bud scoring and aggression prediction in colorectal cancer to enable consistent, reliable and cost-effective patient stratification.
The first chapter of this work explores the feasibility of translating H&E whole slide images to a virtual version of the AE1/AE3 immunohistochemical stain. This tool aims to assist pathologists in tumour bud scoring. We enhanced the loss function of the CycleGAN model to ensure accurate virtual staining while retaining histological structural detail. The modified CycleGAN model demonstrated improved structural similarity and realism in the generated images, providing a proof-of-concept model to create virtual pan-cytokeratin AE1/AE3 whole slide images that could improve the consistency and speed of manual tumour bud scoring.
Secondly, we introduced a method to train an end-to-end tumour bud segmentation system on manual annotations created using reference virtual IHC images. The virtual IHC model highlighted tumour cells or small clusters in the invasive margin of the source H&E slides, resulting in a high-quality dataset of almost 60,000 manually segmented buds. These were used to train a U-Net segmentation model that was evaluated on two distinct patient cohorts. The automated system significantly differentiated between high and low-budding populations, surpassing the manual score’s hazard ratio in one cohort and proving an independent predictor of survival in univariate and multivariate Cox regression. The resulting segmentations also allowed an analysis of the tumour bud areas and distances from the tumour edge, hinting at the possibility of two distinct populations of buds in one cohort.
Finally, self-supervised learning was employed to train a feature extraction network using the VICReg architecture. The encoded representations were clustered, and the model’s success was determined by running Cox regression on the observed probability distribution of clusters across the slides in the patient cohorts. These clusters were correlated with survival and demonstrated the encoding of relevant histological information. To predict tumour aggression, a custom transformer network was then used to analyse the feature vectors from the tumour tissue in the invasive margin. The model demonstrated a stronger correlation with manual budding and provided a more accurate stratification of tumour aggression, improving hazard ratios from the previous model and remaining an independent predictor of survival in univariate and multivariate Cox regression.
This research presents advancements in virtual immunohistochemistry, automated tumour bud scoring and automated tumour aggression prediction in colorectal cancer. We sought to provide a proof of concept for techniques that can pave the way for more consistent, objective, and efficient bud scoring and stratification by aggression. The methods presented, from virtual staining to self-supervised aggression prediction, highlight the synergy between technology and clinical pathology. As we move forward, the fusion of these domains promises to revolutionise diagnostic procedures and usher in an era of more personalised and effective patient care.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Additional Information: Supported by funding from Cancer Research UK.
• Subjects: R Medicine > RC Internal medicine > RC0254 Neoplasms. Tumors. Oncology (including Cancer)
• Colleges/Schools: College of Medical Veterinary and Life Sciences > School of Cancer Sciences
• Supervisor's Name: Insall, Professor Robert and Edwards, Professor Joanne
• Date of Award: 2024
• Embargo Date: 2 February 2026
• Depositing User: Theses Team
• Unique ID: glathesis:2024-84095
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 09 Feb 2024 10:05
• Last Modified: 25 Feb 2026 10:12
• Thesis DOI: 10.5525/gla.thesis.84095
• URI: https://theses.gla.ac.uk/id/eprint/84095