Efficient registration of limited field of view ocular fundus imagery

[摘要] ENGLISH ABSTRACT: Diabetic- and hypertensive retinopathy are two common causes of blindnessthat can be prevented by managing the underlying conditions. Patients sufferingfrom these conditions are encouraged to undergo regular examinationsto monitor the retina for signs of deterioration.For these routine examinations an ophthalmoscope is used. An ophthalmoscopeis a relatively inexpensive device that allows an examiner to directlyobserve the ocular fundus (the interior back wall of the eye that contains theretina). These devices are analog and do not allow the capture of digital imagery.Fundus cameras, on the other hand, are larger devices that o er highquality digital images. They do, however, come at an increased cost and arenot practical for use in theeld.In this thesis the design and implementation of a system that digitises imageryfrom an ophthalmoscope is discussed. The main focus is the developmentof software algorithms to increase the quality of the images to yield results ofa quality closer to that of a fundus camera. The aim is not to match thecapabilities of a fundus camera, but rather to o er a cost-e ective alternativethat delivers su cient quality for use in conducting routine monitoring of theaforementioned conditions.For the digitisation the camera of a mobile phone is proposed. The camerais attached to an ophthalmoscope to record a video of an examination. Softwarealgorithms are then developed to parse the video frames and combinethose that are of better quality. For the parsing a method of rapidly selectingvalid frames based on colour thresholding and spatialltering techniquesare developed. Registration is the process of determining how the selected framest together. Spatial cross-correlation is used to register the frames.Only translational transformations are assumed between frames and the designedalgorithms focuses on estimating this relative translation in a large setof frames. Methods of optimising these operations are also developed. For thecombination of the frames, averaging is used to form a composite image.The results obtained are in the form of enhanced grayscale images of thefundus. These images do not match those captured with fundus cameras interms of quality, but do show a signi cant increase when compared to theindividual frames that they consists of. Collectively a set of video framescan cover a larger region of the fundus than what they do individually. Bycombining these frames an e ective increase in theeld of view is obtained.Due to low light exposure, the individual frames also contain signi cant noise.In the results the noise is reduced through the averaging of several frames thatoverlap at the same location.

[发布机构] Stellenbosch University