Purchase this article with an account.
A Toniappa, SA Barman, AR Fielder, MJ Moseley; Optic Disc Recognition in Images Taken of Premature Infants Using the RetCam 120TM Digital Fundus Camera . Invest. Ophthalmol. Vis. Sci. 2002;43(13):1253.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose:To compare image processing techniques, designed to automatically identify the location of the optic disc on fundus images of premature infants taken with the Retcam 120TM camera system. Methods:Three techniques are described for their reliability in assessing the location of the optic disc. The first two techniques were chosen because of their previously successful application to adult retinal images. The data set used in this analysis of infant fundus images can be divided into two distinct sub-sets, dependent on the ethnicity of the infant; we refer to the two sub-sets as the red set and the blue set. The first method used texture analysis combined with mean averaging over a standard patch size for all images. The second technique used matching correlation, was tested using a number of kernels which were selected on the basis of the two standard image types observed. The third technique of a simple maximum intensity search was included as a comparison to the above established techniques. Results:28 infant fundus images (16 red, 12 blue), all with optic discs, were pre-processed via adaptive, local contrast enhancement. The first technique based on texture analysis gave a specificity of 92% on the blue images and 0% on the red images. The second technique using matching correlation gave a sensitivity of 92% on the blue images and 50% on the red images. The best sensitivity achieved in these experiments occurred when a single kernel selected from the blue images was applied to the entire data set. The third technique that uses a simple intensity search achieved a sensitivity of 67% on the blue images and 6% on the red images. Conclusion:The performance of all three techniques was very dependent on which sub-set of images they were applied to. Overall the texture and correlation techniques work well on the blue sub-set of images. None of the algorithms currently perform to a satisfactory level on the red sub-set of images.
This PDF is available to Subscribers Only