In this thesis I present my work to advance the early detection of oesophageal adenocarcinoma by investigating translational aspects of a minimally invasive oesophageal cell sampling technology for the detection of Barrett oesophagus.

Most oesophageal adenocarcinoma patients present with advanced disease, requiring treatment with chemotherapy with or without radiotherapy, followed by surgery to remove the oesophagus, and even then the overall five-year survival is less than 20%. However, if the cancer can be diagnosed at an early, superficial stage then treatment can be performed endoscopically and over 80% of patients survive beyond 5 years. The disease has a clear pre-cancer stage called Barrett oesophagus, making early detection feasible. A novel test called Cytosponge for diagnosing Barrett by cell collection coupled with an immunohistochemical test (Trefoil factor 3 / TFF3) has been developed.

I have investigated two distinct topics, which are key to implement the Cytosponge-TFF3 test in primary and secondary care. First, I analysed and interpreted data of a pragmatic, prospective, multicentre, randomised controlled trial (BEST3) in order to evaluate the use of Cytosponge in primary care. The study aim was to investigate whether offering this test to patients on medication for gastro-oesophageal reflux disease (GERD) would increase the detection of Barrett oesophagus compared with usual care. We were able to show that in patients with GERD the offer of Cytosponge-TFF3 testing results in improved detection (in excess of 10-fold) of Barrett oesophagus.

Second, I devised and implemented a machine learning framework applied to Cytosponge samples with the objective to reduce the pathologists' screening time. I trained and independently validated the framework on data from two clinical trials, analysing a combined total of 4,662 pathology slides from 2,331 patients. The approach exploits screening patterns of expert gastrointestinal pathologists and established decision pathways to define eight triage classes of varying priority for manual expert review. By substitution of manual review with automated review in low-priority classes, I can reduce pathologist workload by 57% while matching the diagnostic performance of expert pathologists.