Abstract
OBJECTIVE: To describe a new method to estimate the frequency distribution of optic nerve disc area, using digital retinographic images. METHODS AND ANALYSIS: We analysed 492 023 fundus images obtained with seven fundus cameras, mainly in Caucasian subjects. They were grouped by resolution and zoom. They were automatically segmented by identifying the inner edge of the Elschnig scleral ring. For this purpose, a neural network trained by deep learning previously described was used. The number of pixels contained within the segmentation and their frequency distribution were calculated. The results of each camera, using different number of images, were compared with the global results using the Kolmogorov-Smirnov test to confront frequency distributions. RESULTS: The frequency distribution was non-Gaussian, more limited in small sizes than in large ones. If the median is assigned a theoretical value of 1.95 mm(2), the 1th, 5th, 25th, 50th, 75th, 95th and 99th percentiles would correspond to 1.29, 1.46, 1.73, 1.95, 2.20, 2.64 and 3.03 mm(2) in all the dataset. The overall differences were significant for the smaller series, but for each percentile their mean value was only 0.01 mm(2) and the maximum 0.10 mm(2), so they can be considered similar for practical purposes in all cameras. CONCLUSION: By automatically segmenting the edges of the optic nerve and observing the frequency distribution of the number of pixels it delimits, it is possible to estimate the frequency distribution of the disc area in the population as a whole and that of each individual case.