A holistic deep learning framework for endoscopy

Thesis

Endoscopy has developed into a routine method of examination and treatment for the digestive tract. Since most cancer patients will develop precancerous lesions in the early stages of cancer, for example, in colonoscopy, colon polyps are an early stage of colon cancer; in gastroscopy, lesions or ulcers in the mucosa of the stomach are an early sign of gastric cancer. Early endoscopy can detect early cancerous lesions before the onset of obvious symptoms, leading to earlier interventional treatment and reduced patient suffering. Endoscopy can help endoscopists to precisely locate the lesion area by visualising the internal structure of the organ and determine the nature of the lesion based on its shape, size and colour. However, diagnosis by clinicians is usually based on experience and training a specialised physician is time consuming. Therefore, the development of an automated and accurate vision algorithm for endoscopy can assist physicians in their examination to reduce the missed diagnosis of early cancer adenomas and precancerous lesions, thereby avoiding cancer prevention at an early stage and improving the survival rate of patients.

In this thesis, proposed holistic approaches tackle different challenges in endoscopy. Various tasks can assist navigation, treatment planing and minimise missed lesions. Specifically, a novel bounding box pruning approach is proposed to improve artefact detection and an image quality assessment system based on artefact detection is presented. This thesis also proposes new supervised learning and self-supervised learning based approaches for ulcerative colitis grading respectively. Meanwhile, the performance of the proposed self-supervised learning methods in different downstream tasks is explored. In addition, this thesis introduces a spatial-temporal self-attention mechanism with self-supervised learning as an auxiliary task to cope with the challenging problem of video polyp segmentation.

The original contributions are as follows: (1) Pruning-based deep learning artefact detection framework in endoscopy and quality assessment system. (2) Supervised ulcerative colitis grading algorithm using additive angular margin loss. (3) Self-supervised learning with composite pretext-class discrimination for improved generalisability in multi endoscopic image analysis tasks. (4) End-to-end self-supervised and temporal self-attention with feature branching for real-time video polyp segmentation.

Authors: Xu, Z

• DOI: 10.5287/ora-xv1y0rrrn
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Subjects: Medical image analysis