Segmentation of retinal blood vessels using adaptive noise island detection

The blood vessels present in retinal image help in diagnosis and prognosis of several medical pathologies. Automated blood vessel segmentation removes huge amount of efforts and inaccuracies involved with manual segmentation. This paper proposes an unsupervised algorithm for automated vascular tree segmentation. For initial enhancement of the vessel-like structures, our algorithm applies top-hat transform and Difference of Gaussian (DoG) filters. It is observed that various noises in the resultant preprocessed image create some isolated structures: however, the blood vessels remain connected. Based on this observation we propose an adaptive noise island detection (NID) methodology on the preprocessed image to eliminate the noise-like objects while retaining the vascular tree. Our methodology is tested on the publicly available images of DRIVE database. We compared the outcome of the NID algorithm with some recently proposed supervised and unsupervised techniques based on Structural Similarity (SSIM) index.