Development of the system to evaluate soft drusen

Abstract

Abstract: : Purpose: We have developed the system to evaluate the system including the software to identify soft drusens from colour fundus photographs. Methods: Seventeen color fundus photographs of 17 eyes with soft drusens were digitized by a scanner.The digitized images were processed by the newly developed computer analysis software and soft drusens were extracted. This image processing consists of following 5 steps. 1. A fundus image was transformed from color to gray scale data. 2. In order to reduce the influence of an exposure of camera, we removed the bias component by shading correction. 3. A gradient image was acquired by taking the difference between the pixel and the other one which is separated by 5 pixels. This operation was applied to horizontal and perpendicular direction. 4.We acquired candidate pixels of soft druzen by extracting high gradient pixels from above gradient image. Binarization was performed in the region whose center is the candidate pixel. Here, the window size of this region was set to 100 by 100 pixels and the threshold for binarization was set to the average value plus twice standard deviation. 5. Finally, we removed a structure with small area and removed region withlow contrast to eliminate the remained noise. In parallel, soft drusens were identified in colour fundus photographs by an ophthalmologist (the human method), of which the data were used as gold standards to evaluate the computer system. Results: Eight hundred sixty–five soft drusens were detected by the human method, of which 714(82%) soft drusens were extracted by the computer analysis.The defference was statistically signficant(P=.003,Wilcoxon signed–ranks test). One hundred fifty–one soft drusens were not extracted by the computer, of which 68(45%) soft drusens located peri–vascular. Conclusions: Our newly developed computer analysis sytem could detect 82% of the soft drusend detected by ophthalmologist. The drusens located in peri–vasculars were difficult to detect probably becasue those were processed together vasculars images.

Keywords: drusen • image processing • imaging/image analysis: clinical