Development of an automated screening tool for diabetic retinopathy using artificial intelligence

Diabetic retinopathy is the commonest cause of blindness in the working age population in the Western world. It is widely recognised that screening for this treatable condition is highly cost effective. However, there is a shortage in the number of trained personnel required to screen for sight threatening forms of the disease. It has been shown that many of the features of diabetic retinopathy such as microaneurysms, cotton wool spots, exudates and haemorrhages can be identified automatically with high levels of sensitivity and specificity. This work describes the development of an automated computerised system for the screening of diabetic retinopathy through the integration of an artificial intelligent system and the development of custom written software (Diabetic Retinopathy Image Classification Programme) to enable image acquisition, image processing, neural network training and testing to be performed in a structured manner. A combination of conventional image processing and neural network methods are utilised for the identification of the basic features associated with the normal and diabetic fundus image. Preliminary investigations into the identification of sight-threatening features are also described. Identification of normal retinal vasculature and diabetic associated features was performed using three separately trained back-propagtion neural networks. Localisation of the optic disc and macula was achieved by region of interest pixel intensity scanning. Assessment of the optic disc for sight-threatening new vessel growth was performed by comparing the variance in circular intensity profiles of normal optic discs to the variance of those with neovascularisation. Patients were classified as having maculopathy if hard exudates were identified within one disc diameter of the fovea. The overall aim of this project is to develop an automated screening programme for diabetic retinopathy. The initial phase details the development and comparison of a range of algorithms for the detection of features associated with diabetic retinopathy. The final phase details the clinical evaluation of the current screening system