Colorectal Cancer (CRC) is a heterogeneous disease where patient outcomes are often poorly predicted by conventional TNM staging. Quantifying the Tumour Microenvironment (TME), specifically the interaction between tumour and stroma, offers a promising avenue to resolve this unpredictability. However, manual assessment is labour-intensive, and while Deep Learning provides an automated solution, neural networks are prone to failure due to the domain shifts inherent to multi-centre histopathology. This thesis addresses these challenges by investigating the prognostic value of deep learning-derived morphometric biomarkers and developing a framework for human-in-the-loop active learning to enable robust model adaptation.
For the first part of the thesis, a deep learning model was developed to quantify Tumour Cell Density (TCD) across clinically relevant regions of CRC resections. In a multi-centre study (Düsseldorf, n = 127; CLASICC, n = 141), results showed that TCD measured at the Luminal Surface—a region accessible via pre-treatment biopsy—was identified as a robust, independent predictor of cancer-specific survival (HR 2.61, p = 0.010). Comparative analysis suggests that the luminal surface reflects a biologically stable phenotype, validating its utility for pre-operative risk stratification independent of standard clinicopathological factors.
The second half of this thesis addresses the reliability of deployment through Uncertainty Quantification and Active Learning. It investigates the calibration of uncertainty for detecting model error using Monte Carlo Concrete Dropout alongside standard softmax probability and entropy. Results demonstrate that using sparse labels with a proposed post-calibration stage significantly enhances error detection. These metrics were integrated into a novel “offline-online” framework, the “Smart Labelling Tool,” to investigate model adaptation under realistic low-annotation budgets. The study revealed a critical insight that, contrary to standard active learning assumptions, naive Random Subset sampling significantly outperformed Uncertainty-based sampling for model fine-tuning in the low-data regime. While uncertainty sampling failed to accelerate adaptation, it proved superior for Quality Control, identifying nearly twice as many artefacts as random selection. This thesis concludes with evidence that robust clinical AI requires a staged deployment strategy where representativeness precedes refinement.
