Globally 2.2 million people are visually impaired and, of these, approximately 1 million present forms of visual impairment that could be addressed or prevented. Retinal imaging is a key step in the diagnosis and follow-up of major causes of visual impairment. As much as 20% of retinal images collected in the population are affected by artifacts, that render them ungradable both by expert graders and by the more recent automatic grading systems.This work aims to develop an artifact removal strategy able to improve the effectiveness of retinal image grading, in particular for retinal feature segmentation. First, a large group of statistical parameters designed to measure image quality have been selected from the literature. A new ophthalmic database was then collected (CORD - the Comprehensive Ophthalmic Research Database), which includes retinal images with and without artifacts.A mathematical model describing artifacts on the basis of the interaction of the light with the eye during eye photography was then developed. CORD and the mathematical model were then used to train a binary classifier to distinguish pixels affected by distortions within the image without the need for interpretive knowledge of the image itself and, on the basis of this, to establish a validation criterion for quality improvement in retinal images. Finally, an algorithm was developed to isolate in retinal images the regions affected by artifacts, and to subtract from the images the additive contributions to the distortion.The artifact clean-up has been shown to increase the textural information of the retinal images, by improving vessel segmentation by more than 10%. By avoiding the use of interpretative elements of the image, this improvement in the quality of retinal images is agnostic to specific disease processes, and thus potentially applicable to population screening. Further work is necessary to improve the cosmetic quality of the images, to optimise the artifact removal strategy, and to relate the feature extraction improvement to clinical performance.
|Date of Award||14 Dec 2020|
- University Of Strathclyde
|Sponsors||University of Strathclyde|
|Supervisor||Mario Ettore Giardini (Supervisor) & Phil Riches (Supervisor)|