University of British Columbia
Characterization of cellular interactions in histopathology images for clinical and biomedical research using deep learning models
Abstract
dc:descriptionHistopathology remains the gold standard for cancer diagnosis and prognosis, providing detailed insights into tumor architecture and the surrounding microenvironment. Despite recent advances in deep learning, many computational approaches rely on patch-based analysis, which disregards the spatial organization and interactions among individual cells—key components of tumor biology. To overcome this limitation, this dissertation introduces a shift toward cell-based representations by developing a suite of frameworks for cellular graph processing. At the foundation of this approach is a self-supervised model for the identification of cells within the routine histopathology slides. By leveraging the information in the surrounding environment of a cell, this method provides a vector representation for each individual cell without requiring manual annotations. Building upon this, a graph neural network method is developed to characterize the spatial organization of these cells, modeling the patterns in the tumor microenvironment which can be used for predicting the survival time of the patient. These methods are evaluated across diverse cancer types and external datasets, demonstrating a superior performance in survival prediction and a better generalization compared to the state-of- the-art methods in the field. Beyond improving survival prediction metrics, this work explores critical clinical applications. Notably, experiments demonstrate that the proposed method is able to stratify patients with similar genomic profiles (i.e., patient with positive or negative homologous recombination deficiency status) into different cohorts with distinct prognostic trajectories. Furthermore, in an exploratory study, the proposed method demonstrates the capability to predict high-grade serous ovarian cancer response to PARP (poly adenosine diphosphate ribose polymerase) inhibitor therapy using only histopathology images, providing an opportunity for biomarker discovery and treatment selection without requiring for genomic testing. Together, these contributions represent a significant advancement in computational pathology. By integrating self-supervised learning with cellular graph analysis, this dissertation establishes scalable, annotation-free tools for deriving clinically relevant insights from standard histology. These methods bridge the gap between computational innovation and translational oncology, supporting the future of personalized, image-driven cancer care.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy - PhD
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Biomedical Engineering
- Grantor dc:publisher
- University of British Columbia
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Ebrahim Nakhli, Ramin
- Contributors dc:contributor
-
- Bashashati, Ali
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/2429/92466