We used a 600-μm central disc, three annular zones (600–1000, 1000–2000, and 2000–3000 μm diameter), and two outer annular zones (3000–4500 and 4500–6000 μm). The two outer zones were subdivided into four quadrants, giving 8 outer zones, and thus 12 zones in all. The two-threshold method of Otsu
28 was used throughout to define candidate regions in each zone with increased and decreased fluorescence, and local quadratic polynomials were fit to the remaining pixel values, as described in a previously published study.
26 Specifically, the two-threshold Otsu method was applied in each zone to provide an initial segmentation by thresholds k and m into three desired classes: C
0 (nonbackground sources with decreased autofluorescence, e.g., vessels), C
1 (background), and C
2 (areas of increased fluorescence). Because increased FAF was generally in low density, the class C
2 was further subdivided by the one-threshold Otsu method into two classes. The higher pixel values became the new C
2, and the remainder was included in C
1. This method was analogous to the analysis of a low-density drusen image.
26 For each zone, we then had an initial choice of background (C
1) for input to the quadratic polynomial background model. The resultant global model was formed from the 12 local models with appropriate radial and angular cubic spline interpolations at interfaces.
This model of macular background was fit to 10 normal AF images from 10 subjects with normal dilated retinal examinations. The average absolute errors were 3.8% ± 3.5% of net image range. The mean local standard deviations of the original images in each zone (exclusive of the hypofluorescent and hyperfluorescent pixels) ranged from 3.0% to 4.1% over the 10 images. If these mean local standard deviations are taken as representative of noise in the image, it follows that the errors of the model were of the same magnitude as the noise in the original data. Each AF image was then leveled by subtracting its background model with an offset of 125 gray levels, and the mean and SD σ of the leveled image (excluding vessels) was calculated. We found that the leveled image fell within 2.0 σ of the mean for 99.7% of pixels in each of the images
(Fig. 1) . (By contrast, if the gray levels of the image had a normal distribution, then gray levels above 2.0 σ would comprise 2.3% of the image.) We therefore defined increased FAF in this study as a gray level greater than 2.0 σ above the image mean, after the image has been leveled by the model.