Purchase this article with an account.
Daniel Albertus, Kanishka Thiran Jayasundera, Chris Ranella, Matthew Johnson-Roberson; Automated Detection of the Optic Disc in Fundus Autofluorescence Images of Age-Related Macular Degeneration. Invest. Ophthalmol. Vis. Sci. 2016;57(12):1626.
Download citation file:
© 2017 Association for Research in Vision and Ophthalmology.
Current subjective analysis of fundus autofluorescence (FAF) images for age-related macular degeneration (AMD) is not ideal given new therapeutics which may be able to slow the progression of dry AMD. The detection of both hyper- and hypofluorescent regions of disease is challenging due to similarities between these areas and various normal FAF features including the optic disc. We show a reliable way of identifying the optic disc in a fully automatic manner in images from patients with AMD.
Fundus autofluorescence (FAF) images from the Heidelberg Spectralis were first pre-processed, which includes correction for uneven illumination and contrast, followed by segmentation of the blood vessels. Next, a pixel-wise classification using Gabor wavelet responses and pixel intensity as features was carried out, followed by spatial clustering of the positively classified pixels to obtain distinct regions of interest. Region-wise classification was then carried out on these distinct clusters and the optic disc was identified based on surrounding vessel density and orientation as well as circle fit and relative size.
Optic disc detection was performed on 144 FAF images of patients with AMD. Of these 144 FAF images, 29 images were used for training whereby a human grader determined where the optic disc was on the image. The remaining 115 images were then used for testing of the algorithm which was successful in correctly identifying 86 (75%). Pixel wise the algorithm had a sensitivity of 97% and a specificity of 99% indicating that the algorithm was able to not only identify where the optic disc was but also accurately identify the specific pixels which represented it compared to the human grader.
We present a method to automatically detect and segment optic discs in FAF images of patients with dry AMD. The proposed method takes advantage of vessel geometry outside the optic disc along with the geometry of the optic disc itself to refine positive pixel classifications. Results validated the effectiveness of the technique on actual clinical FAF images despite the presence of geographic atrophy which mimics the appearance of the optic disc.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.
The process leading up to optic disc detection including vessel segmentation, pixel-wise classification and circle fitting.
Examples of the optic disc being positively identified despite the similar appearance to geographic atrophy.
This PDF is available to Subscribers Only