Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.797567
Title: A retinal vasculature tracking system guided by a deep architecture
Author: Uslu, Fatmatülzehra
ISNI:       0000 0004 8504 4407
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2017
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
Many diseases such as diabetic retinopathy (DR) and cardiovascular diseases show their early signs on retinal vasculature. Analysing the vasculature in fundus images may provide a tool for ophthalmologists to diagnose eye-related diseases and to monitor their progression. These analyses may also facilitate the discovery of new relations between changes on retinal vasculature and the existence or progression of related diseases or to validate present relations. In this thesis, a data driven method, namely a Translational Deep Belief Net (a TDBN), is adapted to vasculature segmentation. The segmentation performance of the TDBN on low resolution images was found to be comparable to that of the best-performing methods. Later, this network is used for the implementation of super-resolution for the segmentation of high resolution images. This approach provided an acceleration during segmentation, which relates to down-sampling ratio of an input fundus image. Finally, the TDBN is extended for the generation of probability maps for the existence of vessel parts, namely vessel interior, centreline, boundary and crossing/bifurcation patterns in centrelines. These probability maps are used to guide a probabilistic vasculature tracking system. Although segmentation can provide vasculature existence in a fundus image, it does not give quantifiable measures for vasculature. The latter has more practical value in medical clinics. In the second half of the thesis, a retinal vasculature tracking system is presented. This system uses Particle Filters to describe vessel morphology and topology. Apart from previous studies, the guidance for tracking is provided with the combination of probability maps generated by the TDBN. The experiments on a publicly available dataset, REVIEW, showed that the consistency of vessel widths predicted by the proposed method was better than that obtained from observers. Moreover, very noisy and low contrast vessel boundaries, which were hardly identifiable to the naked eye, were accurately estimated by the proposed tracking system. Also, bifurcation/crossing locations during the course of tracking were detected almost completely. Considering these promising initial results, future work involves analysing the performance of the tracking system on automatic detection of complete vessel networks in fundus images.
Supervisor: Bharath, Anil Anthony Sponsor: Ministry of National Education, Turkey
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.797567  DOI:
Share: