Page 94 - Computational Retinal Image Analysis
P. 94

86     CHAPTER 5  Automatic landmark detection in fundus photography




                            The best results for this method occur when the fuzzy model is calculated on
                         the illumination corrected image using a multiscale approach to vessel segmenta-
                         tion (vessel segmentations that target different widths of vessels in the image). The
                         method is run on six different scales of vessel segmentations. If there is no consensus
                         on the vessel voting, the largest bright region from a predetermined number of pixels
                         in the illumination corrected image is used as the found OD. The method was tested
                         on the STARE database (81 images) and achieved 89% accuracy (100% accuracy on
                         the 31 healthy retinas).


                         5.3  Detection of optic disc in retinal images by means of a
                         geometrical model of vessel structure (Foracchia et al., 2004 [21])
                         This method is noted both for being state-of-the-art at its time of publication and
                         beyond, but also for its ability to detect the OD even when it was not present in the
                         image. This method is similar to the method developed by Hoover and Goldbaum
                         (2003) by which they both are looking for the convergence of vessels at the OD. This
                         method, however, creates a geometric model that is fit to the directional information
                         of all the blood vessels in the image. This model, which is essentially two parabolic
                         arches, finds the OD center as the convergence of these two parabolas, symbolizing
                         the convergence of the retinal vasculature.
                            To develop the model, an arbitrary vessel tracking algorithm can be used to pro-
                         vide vessel centerline position and direction. The two parabolas, which come to-
                         gether at the OD and represent the main vessel paths across the retina can be describe
                         by the following locus of points:
                                                   Γ = ( ) ay 2  = }                      (3)
                                                      { xy
                                                        ,
                                                                x
                                                           :
                            Where a controls the aperture of the parabola. This only describes the course of
                         the main vessels in the image. To capture the fact that the vessels branch and bifur-
                         cate, with vessels inside the parabola bending toward the center of the image and
                         vessels outside the parabola bending toward the periphery. To account for this, the
                         complete model for vessel direction in the image is:

                                                                                        (4)
                                                     
                                      θ mod ( xy,,p) = arctan sgn x () sgn y ()  1  + (  
                                                     
                                                                        dx y,,p)
                                                               2  ax        
                                                                              
                            Where
                                                         y − sgn  y ()  x
                                                  (
                                                 dx y,,p) =        a                      (5)
                                                           c 1  +  c 2
                                                         1 +  e − x  1 +  e x
                            p contains the model parameters a and the optic disc center points. In a cartesian
                         coordinate system, the x and y values in these formulas should be shifted by values of
   89   90   91   92   93   94   95   96   97   98   99