Automated retinal image matching by segmentation

Abstract

Spectral characterization of retinal images is a complex task because of the superimposed blood vessels in the retinal photographs. We have developed a segmentation technique based on a combination of edge detection and optimal thresholding operations so that the blood vessel structure can be removed from the retinal texture. The blood vessel removal method includes edge detection, tracing, and replacement with gray levels closely approximating the behavior of the background. The edge detection uses the (unsigned) first derivative of the image. The tracing routine is controlled by a mouse and follows the crests of the first derivative to trace out a length of edge. When the blood vessels are outlined, a search pattern based on finding a dark feature of given width in a light background locates those pixel elements that are clearly blood vessels. When the blood vessels are removed, they are replaced by a close approximation of the background. After the end point grey levels are determined, the vessel is painted over with a linear replacement operation, matching the background at both sides of the vessel. Spectral characterization and matching of the retinal texture can then be performed for automated detection of textural changes.