Texture analysis of retinal neovascularisation

The symptoms of diabetic retinopathy can be visualised or photographed directly through the pupil using a fundus camera, but the structural changes or lesions are indistinct, too numerous or easily confused with normal retinal features. Our approach to the detection and quantification of the features is based on the classical model of computer vision, involving acquisition of the image, pre-processing, segmentation and classifications. Quantification is a trivial step at the end of this process. This paper describes our investigation into the segmentation and classification stages of this task, and we present the performance results of the developed system. We have investigated image texture analysis as a method of giving an objective measure to subjective quantities such as 'roughness' and 'smoothness' of the image. We shall show that the lesions which occur in DR, while showing only slight changes from the rest of the image in terms of intensity information, demonstrate radically different texture appearances. We first describe the acquisition and processing of our images, and the image texture measures applied. We then show that statistical differences exist between different image features, and describe our methodology for segmenting and classifying an entire retinal image based upon texture.