Maximum Likelihood Retinal Blood Vessel Detection by Scale–Space Filtering

Abstract

Abstract: : Purpose: The detection of retinal blood vessels and abnormalities in retinal images is of interest for diagnosis of disease, particularly diabetes. Automatic detectors are of interest as the first stage of an automatic analysis system for large–scale screening programs. Initially, this system is to detect only blood vessels. Methods: Blood vessels in different parts of the retina look very similar, differing only in their size and orientation. Scale–space methods are designed to recognise similar structures at different scales by applying Gaussian derivative filters of different sizes, and looking for the size–independent characteristics of vessels. A maximum likelihood based method is used to estimate a score parameter for classification of the image into background and vessel regions. Noise is accounted for in the forward model. Results: Manual segmentation of blood vessels was used as a reference standard. Using an approximation to the (complicated) maximum likelihood inversion, the detector is shown to have a sensitivity of 68% and a specificity of 95% in classifying vessel pixels in a thirty degree retinal image. Conclusions: Although the system uses an approximation to the maximum likelihood method, it has been shown to be an accurate method of pixel classification. The result is a continuous measure of the likelihood of a pixel being in a vessel or not, and avoids the need for region growing or vessel tracking techniques.

Keywords: image processing • diabetic retinopathy