Automated retinal analysis

Abstract

Diabetes is a chronic disease affecting over 2% of the population in the UK [1]. Long-term complications of diabetes can affect many different systems of the body including the retina of the eye. In the retina, diabetes can lead to a disease called diabetic retinopathy, one of the leading causes of blindness in the working population of industrialised countries. The risk of visual loss from diabetic retinopathy can be reduced if treatment is given at the onset of sight-threatening retinopathy. To detect early indicators of the disease, the UK National Screening Committee have recommended that diabetic patients should receive annual screening by digital colour fundal photography [2]. Manually grading retinal images is a subjective and costly process requiring highly skilled staff. This thesis describes an automated diagnostic system based oil image processing and neural network techniques, which analyses digital fundus images so that early signs of sight threatening retinopathy can be identified. Within retinal analysis this research has concentrated on the development of four algorithms: optic nerve head segmentation, lesion segmentation, image quality assessment and vessel width measurements. This research amalgamated these four algorithms with two existing techniques to form an integrated diagnostic system. The diagnostic system when used as a 'pre-filtering' tool successfully reduced the number of images requiring human grading by 74.3%: this was achieved by identifying and excluding images without sight threatening maculopathy from manual screening.