Model based segmentation for retinal fundus images

This paper presents a method for detecting and measuring the vascular structures of retinal images. Features are modelled as a superposition of Gaussian functions in a local region. The parameters i.e. centroid, orientation, width of the feature are derived by a minimum mean square error (MMSE) type of spatial regression. We employ a penalised likelihood test, the Akakie Information Criteria (AIC), to select the best model and scale for vessel segments. A maximum-cost spanning tree (MST) algorithm is then used to perform the neighbourhood linking and infer the global vascular structure.. We present results of evaluations on a set of twenty digital fundus retinal images.