Purchase this article with an account.
G. Troglio, M. Alberti, J. A. Benediksson, G. Moser, S. B. Serpico, E. Stefansson; Unsupervised Detection of Temporal Changes in Fundus Images. Invest. Ophthalmol. Vis. Sci. 2010;51(13):3858.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
This research contributes to the development of a system able to detect structural changes in multitemporal images of the human retina, associated with retinopathy or systemic diseases.
The proposed approach is based on the application of unsupervised change detection techniques to pairs of images, taken of the same patient at different times. First, the images need to be registered: This is achieved in an automatic way, by using a genetic optimization technique. Then, the registered images are compared, in order to generate two difference images obtained by a pixel-by pixel subtraction of the first date image from the second date one, and viceversa. Finally, a multiple classifier approach is used in order to detect the temporal changes: In particular, a thresholding technique is applied to a set of randomly distributed sub-images of the image difference. For each window a change sub-map is obtained. Subsequently, for each pixel of the image, a fusion of label outputs is performed, and a global change map is obtained.Images were acquired from Icelandic patients attending a retinopathy screening service, by a Canon CR6-45NM fundus camera.
The algorithm was tested on eight pairs of color fundus images, which were taken during different medical visits. The change map, obtained by the proposed approach, and a test map, generated by a human observer, were compared. A sensitivity value about 80% on the average, with standard deviation about 17, and specificity about 98% on the average, with standard deviation about 1.2, were obtained. The figure shows results for a pair of color images: The change map (a) and the test map (b) superimposed to one of the two images to be compared (color legend: white = new white spots, red = new red spots, blue = old white spots, green = old red spots).
A quantitative assessment of the change detection performances suggests that the proposed method is able to provide accurate change maps. The results are good and in accordance with the performance specifications recommended by the British Retinopathy guidelines.
This PDF is available to Subscribers Only