We have presented an algorithm to segment retinal blood vessels from both red-free and fluorescein images which combines: 1) the multiscale representation that gives information about the width of blood vessels (it does not metter which orientation they might have in the image), 2) two geometrical properties of tube-like structures based on the first and second derivative information which give weights to pixels with a high probability of belonging to vessels, and 3) a multiple pass region growing procedure. The region growing algorithm is relatively fast because in the initial stage growth is restricted to regions with low gradients, allowing vessels to grow where the values of the maximum principal curvature lie within a wide interval. This allows rapid growth outside the boundary regions. In the final stage when the borders between classes are defined, the algorithm grows vessel and background classes simultaneously without the gradient restriction.
It should be pointed out that although the idea of the diameter-dependent equalisation factor applied to the multiscale information is empirical, it can be justified by a simple model in which the intensity of the image is proportional to the amount of blood in the light path corresponding to that pixel. In any case, the algorithm seems to work robustly in detecting retinal blood vessels over a large range of widths.
The first validation of measurements, for diameters and branching angles, showed a significant difference between manual and automatic measurements. These significant differences can not be taken as error but only as indications of the variation between different measurement techniques. From the comparison of automatic measurements between red-free and fluorescein paired images we did not find any significant difference for either diameters or branching angles. Because of the better quality of fluorescein images, measurements taken from this technique are normally considered to be more accurate than those from red-free, and are taken as a gold standard measure [7,9]. However, it has been reported [5,26] that measurements of vessel diameters taken from fluorescein angiographs are larger than red-free measurements because the higher concentration of fluorescein dye in the cell-free layer near the vessel wall. We conclude from our comparisons that the approach presented gives comparable estimates of diameter and branching angles in both image types but the accuracy of these measurements remains to be evaluated.
A similar comparison between manual measurements of red-free and fluorescein images showed a significant difference in the vessel diameter measurements, fluorecein being larger than red-free. However, the Bland-Altman plot (Figure 14(b)) showed a tendency for fluorescein diameters to be generally smaller than red-free for small vessel widths, whereas they are larger for larger vessel widths. This bias is not observed in the automatic measurements (Figure 14(a)).
The appearance of the retinal blood vessels can be an important diagnostic indicator of various clinical disorders of the eye and the body. Quantitative analysis of retinal blood vessels from fundus images have usually been studied in terms of individual bifurcations, measuring a few of the most clearly visible bifurcations in an image. Accurate detection of vessels widths and continuous blood vessels networks, based on a robust segmentation, will allow the study of morphological changes resulting from disease: geometrical factors such as diameters, branching angles, lengths, tortuosity, length to diameter ratios, etc. and network properties such as connectivity, branch ordering, purely topological indices which describe the shape of branching trees, and, if appropriate, fractal properties.