This section describes methods for modeling the appearance of retinal vessels in fundus images in a manner that is sufficient for the purpose of image segmentation. Specifically, we are interested in features of vessels that allow an algorithm to determine if specific pixels in an image are part of the vessel tree or the background. A full physics-based modeling of the image formation process, including the geometry of the vasculature and the imaging systems, could yield more accurate and realistic models, but is beyond the scope of this chapter (see [16, 35,89,100]).
Generally, models are applicable only in the context of the image type for which they are developed. For example, in a red-free retinal image (that is, an image taken with a specialized red filter to block out most of the red light), the blood vessels appear to be darker than the local background. In a fluorescein image, dye is injected into the bloodstream, which causes the vessels to appear brighter than the background. Application of the same intensity-based model in both cases will not yield the same results. Fortunately, it is often possible to modify a model that works well at detecting light vessels against a dark background to be able to detect dark vessels against a light background.
Most of the methods described in the literature model the cross-sectional profile of vessels. Figure 6.7 illustrates the idea of a cross section. The conceptual basis for such modeling is the differential behavior of light propagation through vessels, and the reflectance off of the vessel surface, compared to the local background. The following sections review some of the different models used to detect retinal vessels and assume an image in which blood vessels are darker than the surrounding background.
Was this article helpful?