Page 96 - Computational Retinal Image Analysis

88 CHAPTER 5 Automatic landmark detection in fundus photography

vertical dark bar and different width X patterns running through the middle of the
circle. While the dark bar and X patterns make sense in trying to represent the darker
vasculature, the bright circle on black background provided the best results through
empirical means. This method also adapts the template to the pixel size of the OD
based on the field of view (FOV) of the image, pixel footprint and known average
values for retinal area and OD size in millimeters. The formulas to calculate the OD
size in pixels then become:
A
f img = FOV (6)
N FOV
where the image foot print, f img , is equal to A FOV , a known area that a certain FOV
covers, over N FOV , the pixel footprint of the image. For instance, a 45° FOV has an
2
average area of 124.8 mm . The radius of the OD in pixels, r OD_img , then becomes:
D ( / ) 2 2
r = OD (7)
_
OD img
f img
Where D OD is a known average value for the disc diameter, in this case 1.85 mm.
Using these formulas, the template can be adjusted based on the pixel footprint and
FOV to help ensure accurate template matching.
The actual template matching is performed on an illumination corrected version
of the image. To speed up the matching, the template is correlated (using the Pearson
correlation coefficient, c ij ) to the image on a grid, as opposed to each pixel in the
image.
( (
∑ fx y) − ) ( ( tx iy − j) − )
−
,
,
t
f
c = xy , m m (8)
ij
( (
( (
2
2
(∑ xy , fx y) − ) ) ∑ , ( xy tx − iy, − ) − ) )
−
t
j
f
,
m
m
where t m and f m are the mean intensities of the template and the sub-image overlap-
ping the template. This speeds up the processing, recognizing that this is the first of
a two-step process that will be refined. To avoid being stuck in a local maximum that
may or may not be the OD (due to bright lesions or camera artifacts), a small percent-
age of candidates are chosen for the second step of processing.
Once the candidates have been found, the final OD is localized through vertical
matched filtering. The main vessel arcades leave the OD vertically before curving to
a parabolic shape around the retina. A Gaussian kernel is used to match the intensity
profile across the vessels in the green channel image through convolution.
− x 2 L
, (
Gx y) =− ae 2σ 2 , for y ≤ (9)
2
The length, L, is the length for which the vessel has fixed orientation. The size of the
Gaussian kernel should be fixed to the width of the main retinal vessels. The candidate