Robust segmentation of vessels from retinal angiography

This paper presents an algorithm based on mathematical morphology and linear processing for vessels recognition in a noisy retinal angiography. A geometrical model of undesirable patterns is defined in order to separate the vessels from their environment. Bright round peaks are first extracted, allowing segmentation of microaneurisms from suitable angiograms of diabetic patients. Next, linear structures are extracted using mathematical morphology, then appropriate differential properties are computed using a Laplacian filter. Finally, vessels are extracted using curvature differentiation. Results on various medical data from a normal eye and from a set of different abnormalities are presented and show that this algorithm can be used as a robust segmentation tool. We intend to use it as a first step for image registration.