University of British Columbia
Task-dependent utility of computational pathology in colorectal cancer
Abstract
dc:descriptionColorectal cancer (CRC) is biologically heterogeneous, creating demand for molecular and prognostic tests that can guide clinical care. Computational pathology may recover clinically relevant information from routinely collected haematoxylin and eosin-stained whole-slide images, but its utility depends on label reliability, generalizability, and the threshold at which predictions are used. This thesis evaluated consensus molecular subtype (CMS), microsatellite instability (MSI), KRAS/NRAS hotspot mutations, BRAF-V600E, and overall survival across The Cancer Genome Atlas (TCGA), Personalized OncoGenomics (POG), and a MD Anderson Cancer Center (MDA) cohort. Four histopathology foundation models, with weights frozen, were leveraged for attention-based multiple-instance learning (MIL). RNA-based CMS analysis showed that different classification procedures gave associated, but not highly concordant label outputs. The multi-sample, random-forest CMS classifier (RF-CMS), and centroid-based, single-sample predictor (SSP-CMS) had limited agreement when indeterminate calls were retained (kappa = 0.28), improved among jointly classified cases (kappa = 0.56). Correlative analyses performed to measure biological coherence of classifier outputs were strongest for CMS1 immune and CMS4 stromal programmes. Histopathology-based CMS prediction using RF-CMS as a ground truth consequently showed limited performance. MSI prediction achieved 0.944 sensitivity, 0.501 specificity, and 0.983 negative predictive value in held out evaluation. For illustration of use-case, a cohort of 1000 patients was simulated: 43.9% of patients were assigned to the low probability of MSI group, with a false positive rate of 1.8%. KRAS/NRAS hotspot mutation rule-out sensitivity reached 0.93-0.99 only at specificity of 0.06-0.13, whereas BRAF-V600E mutation rule-in specificity reached 0.96-0.97 at sensitivity of 0.12. Survival models based on UNI and prov-GigaPath separated the pooled test set, but differing survival reference points and limited within cohort discrimination prevent a clean interpretation. These findings demonstrate that computational pathology should be evaluated as an endpoint specific measurement rather than as a uniformly transferable technology. Reliable labels, development-derived thresholds, external validation, cohort-stratified analysis, and explicit clinical workflows are necessary before image derived predictions can support prospective care.
Degree
thesis:*- Name thesis:degree_name
- Master of Applied Science - MASc
- Level thesis:degree_level
- master's
- Discipline thesis:degree_discipline
- Biomedical Engineering
- Grantor dc:publisher
- University of British Columbia
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Diaz-Stewart, Michael
- Contributors dc:contributor
-
- Bashashati, Ali
- Loree, Jonathan M.
Rights
dc:rights- Statement dc:rights
-
- Attribution-NonCommercial-NoDerivatives 4.0 International
- Language dc:language
- eng
Identifiers
dc:identifier.*- Handle dc:identifier
- http://hdl.handle.net/2429/95830