Fourier cross-sectional profile for vessel detection on retinal images

Retinal blood vessels are important objects in ophthalmologic images. In spite of many attempts for vessel detection, it appears that existing methodologies are based on edge detection or modeling of vessel cross-sectional profiles in intensity. The application of these methodologies is hampered by the presence of a wide range of retinal vessels. In this paper we define a universal representation for upward and downward vessel cross-sectional profiles with varying boundary sharpness. This expression is used to define a new scheme of vessel detection based on symmetry and asymmetry in the Fourier domain. Phase congruency is utilized for measuring symmetry and asymmetry so that our scheme is invariant to vessel brightness variations. We have performed experiments on fluorescein images and color fundus images to show the efficiency of the proposed algorithm technique. We also have performed a width measurement study, using an optimal medial axis skeletonization scheme as a post-processing step, to compare the technique with the generalized Gaussian profile modeling. The new algorithm technique is promising for automated vessel detection where optimizing profile models is difficult and preserving vessel width information is necessary.