Use this URL to cite or link to this record in EThOS:
Title: Vision-based retargeting for endoscopic navigation
Author: Ye, Menglong
ISNI:       0000 0004 6422 7810
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2017
Availability of Full Text:
Access from EThOS:
Access from Institution:
Endoscopy is a standard procedure for visualising the human gastrointestinal tract. With the advances in biophotonics, imaging techniques such as narrow band imaging, confocal laser endomicroscopy, and optical coherence tomography can be combined with normal endoscopy for assisting the early diagnosis of diseases, such as cancer. In the past decade, optical biopsy has emerged to be an effective tool for tissue analysis, allowing in vivo and in situ assessment of pathological sites with real-time feature-enhanced microscopic images. However, the non-invasive nature of optical biopsy leads to an intra-examination retargeting problem, which is associated with the difficulty of re-localising a biopsied site consistently throughout the whole examination. In addition to intra-examination retargeting, retargeting of a pathological site is even more challenging across examinations, due to tissue deformation and changing tissue morphologies and appearances. The purpose of this thesis is to address both the intra- and inter-examination retargeting problems associated with optical biopsy. We propose a novel vision-based framework for intra-examination retargeting. The proposed framework is based on combining visual tracking and detection with online learning of the appearance of the biopsied site. Furthermore, a novel cascaded detection approach based on random forests and structured support vector machines is developed to achieve efficient retargeting. To cater for reliable inter-examination retargeting, the solution provided in this thesis is achieved by solving an image retrieval problem, for which an online scene association approach is proposed to summarise an endoscopic video collected in the first examination into distinctive scenes. A hashing-based approach is then used to learn the intrinsic representations of these scenes, such that retargeting can be achieved in subsequent examinations by retrieving the relevant images using the learnt representations. For performance evaluation of the proposed frameworks, extensive phantom, ex vivo and in vivo experiments have been conducted, with results demonstrating the robustness and potential clinical values of the methods proposed.
Supervisor: Yang, Guang-Zhong ; Teare, Julian Sponsor: Imperial College London
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral