Gastrointestinal cancer, including colorectal and oesophageal cancer, accounts for over 35% of cancer-related deaths worldwide. It is possible to identify these diseases at an early stage during endoscopic examinations, as well as facilitate prompt treatment to enhance patient outcomes.

For colorectal cancer, deep-learning approaches have shown increases in polyp detection rates during colonoscopies. However, most of these systems are trained using static images, whilst, in clinical practice, the procedure is conducted on a real-time video feed. Moreover, enhanced polyp detection rates may result in the identification of benign polyps, leading to unnecessary increased time and cost. Consequently, there is a growing demand for accompanying tools to characterize polyps in order to determine which ones require resection. Recent optical diagnosis deep-learning approaches have shown promising results assisting with this task, yet polyp appearance during a procedure can vary, making automatic predictions
unstable.

Patients with Barrett’s esophagus, a recognized precursor to adenocarcinoma in esophageal cancer, undergo gastroscopies to diagnose and treat early dysplasia. Studies indicate that up to 25% of early cases are missed during gastroscopies. Although deep-learning approaches have been investigated as decision-support tools, the complexity of the task hinders their translation to clinical practice; the lesions
are often subtle, evading human eye detection, and the data presents logistical difficulties as it comprises lengthy videos with limited lesion variation.

This thesis explores deep-learning methods to bridge the gap between the development and translation of computer-aided detection and diagnosis tools for early cancer detection in endoscopy. In colonoscopies, temporal information and features from videos are harnessed to solve low-stability challenges and improve performance in clinical scenarios. In gastroscopies, prior knowledge is utilised during data preparation and model training to reduce overfitting and obtain more generalisable solutions.