B-COSFIRE and Background Normalisation for Efficient Segmentation of Retinal Vessels

Abstract

Ocular diseases are becoming more common these days, especially the ones directly related to diabetes such as diabetic retinopathy (DR). This can be attributed to the steady increase in the number of diabetic cases worldwide. DR can be treated if diagnosed early enough to prevent further damages to retina which could lead to total blindness if left untreated. Automatic segmentation of retinal blood vessels from fundus image is useful towards development of an efficient and accurate computer-assisted retinal diagnosis system. In this study, we propose a combination of B-COSFIRE filter and background normalization methods to segment the retinal blood vessels (RBVs) from fundus image. By performing background normalization on B-COSFIRE filter output and combining it with the original B-COSFIRE output, the segmentation performance can be improved in terms of Sensitivity. Validation of the proposed method on two public databases, namely DRIVE and STARE shows comparable segmentation performance to published methods with improved Sensitivity. The method achieves Sensitivity values of 78.33% and 81.04% and Specificity values of 96.51% and 95.69% for DRIVE and STARE database, respectively.