Automated extraction of retinal vasculature

Purpose: The authors propose an algorithm that automatically extracts retinal vasculature and provides a simple measure to correct the extraction. The output of the method is a network of salient points, and blood vessels are drawn by connecting the salient points using a centripetal parameterized Catmull–Rom spline. Methods: The algorithm starts by background correction. The corrected image is filtered with a bank of Gabor kernels, and the responses are consolidated to form a maximal image. After that, the maximal image is thinned to get a network of 1-pixel lines, analyzed and pruned to locate forks and form branches. Finally, the Ramer–Douglas–Peucker algorithm is used to determine salient points. When extraction is not satisfactory, the user simply shifts the salient points to edit the segmentation. Results: On average, the authors’ extractions cover 93% of ground truths (on the Drive database). Conclusions: By expressing retinal vasculature as a series of connected points, the proposed algorithm not only provides a means to edit segmentation but also gives knowledge of the shape of the blood vessels and their connections.