Deep Learning in Oesophageal Cancer Development: Integrating Multi-Stain Histopathology Images of the Capsule Sponge

Barrett’s oesophagus (BE) is a premalignant condition that increases the risk of oesophageal adenocarcinoma (OAC). Surveillance relies on endoscopy with random biopsies, a costly and invasive procedure that limits population-scale feasibility. The capsule sponge, a pan- oesophageal cell collection device, has emerged as a minimally invasive alternative, but introduces unique computational challenges: slides are far larger than biopsies, cellular context is often absent, empty regions dominate, and immunohistochemistry (IHC) slides contain control tissue that complicates automated analysis. This thesis develops and evaluates a computational pipeline to address these challenges and enable scalable risk stratification for BE. First, I present non-AI based pre-processing techniques, including a novel tissue seg- mentation framework (TissueTector), colour filtering algorithms for IHC slides, and robust registration via the HistoKatFusion framework. These methods substantially reduce computa- tional burden, with colour filtering achieving a 20-fold runtime reduction while maintaining clinical relevance. Validation across multiple datasets and staining protocols confirms robust- ness and generalisability. Second, I introduce AI-based models for gland detection and classification within a pathologist-in-the-loop workflow. Using transfer learning with YOLOv11, I trained seg- mentation models for H&E and IHC images, followed by classification models for p53 (positive, equivocal, negative), TFF3 (positive, negative), and H&E glands (normal, atypical, dysplastic). Trained on enriched subsets of the DELTA and BEST2 trials and validated on independent cohorts, these models achieved high sensitivity with zero false negatives, ensuring clinically relevant cases are not missed. The atypia model, however, performed poorly. Finally, Shiv Sakthivel and I developed an interpretable algorithm that detects atypia based on glandular and nuclear features already used in pathology practice. Together, these methods establish a scalable, efficient, and clinically applicable framework for AI-assisted diagnosis in BE, paving the way for earlier detection and improved outcomes in population- level screening and patient-level surveillance.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2025

Author dc:creator

• Markert, Greta

Advisors dc:contributor.advisor

• Markowetz, Florian
• Fitzgerald, Rebecca

• histopathology
• barrett's esophagus
• AI
• computer vision

Licence

• https://www.repository.cam.ac.uk/bitstreams/0eedd68a-1bff-4a90-8b83-fe5ccd96153e/download
• http://purl.org/NET/rdflicense/allrightsreserved

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.125950

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/396657