Abstract
Abstract: :
Purpose:To demonstrate accurate models of normal macular images that could form the basis for digital analysis of pathology in three key macular imaging modalities. Methods:We studied standard fundus photographs; autofluorescence (AF) and infrared (IR) images were obtained on the Heidelberg HRA. We have previously presented models of foveal photographic reflectance and AF based on quadratic polynomials. Here we extended these models of normal background to the entire macula (6000– µ diameter region divided into 27 zones, including significant retinal vasculature at the arcades. Each step was completely automated. Initial classification in each zone into hyporeflectant (hypofluorescent), background and hyperreflectant (hyperfluorescent) classes C0, C1 and C2 was chosen by the two threshold Otsu method. The background input data C1 was fit by a quadratic Q, and the image data Z was leveled by Z1 = Z – Q + 125. The collection of quadratics in each zone formed the background model for the image. The local standard deviation s and local median M of the leveled image Z1 were calculated and final segmentation into C0, C1 and C2 was done in each zone in terms of multiples of s above or below M. We used the local criteria 2.5s above M for hyperAF/IR and 2s above M for photographic hyperreflectance. We also calculated the local standard deviation of the original image in each zone as an estimate of noise in the original data. Results:. The average absolute errors of the model fit to 10 normal photographic, AF and IR images exclusive of vasculature were 6.4 +/– 7.3%, 3.6 +/– 3.4%, and 4.6 +/– 4.5%, respectively, of net image range. The class C2 of hyperreflectant (fluorescent) pixels was .01% for the fundus photographs, 0.089% of the area in the AF images and 0.38% of the area in the IR images The mean local standard deviations of the original images over all zones ranged from 5.5 to 7.0%, 3.0 to 4.1%, and 4.0 to 5.1% respectively. Conclusions: These unified models of normal AF, IR and photographic images had an accuracy that was essentially only limited by the magnitude of the noise in the original data. The characterizations of normal image variability obtained from the models could thus be used to define objective criteria for autofluorescence, infrared and photographic abnormalities and allow objective analysis of macular disease processes.
Keywords: imaging/image analysis: clinical