Purchase this article with an account.
Katia Delalibera Pacheco, Yulia Wolfson, Philippe Burlina, David E. Freund, Albert Feeny, Neil Joshi, Neil M Bressler; Evaluation of automated drusen detection system for fundus photographs of patients with age-related macular degeneration. Invest. Ophthalmol. Vis. Sci. 2016;57(12):1611.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Evaluate an automated grading system for classifying age-related macular degeneration (AMD) severity from fundus photos using a novel method developed at Johns Hopkins compared with human graders.
Retinal images were analyzed to try to improve detection of people with earlier stages of AMD among individuals over age 50 who should be referred to eye care providers to monitor for onset of choroidal neovascularization (CNV). Using already digitized and de-identified images from the National Institutes of Health (NIH) database of fundus images from the Age-Related Eye Disease Study (AREDS) dbGaP dataset, including images graded using 4 categories of: (1) no AMD, (2) early, (3) intermediate, or (4) advanced stage of AMD, per AREDS stages. In addition, differences between two operators grading and differences between manual grading (operator) and automated grading (machine) for the same set of images were compared. A two class problem was used; class 1 (AMD category 1 or 2) vs. class 2 (AMD category 3 or 4). Two independent graders reviewed and classified all images into these specific categories. Paired gradings were compared and percentage agreement using kappa statistics were calculated.
A subset of 917 images were evaluated. Agreement was almost perfect when comparing the first grader and AREDS (kappa = 0.87). For the second grader and AREDS, agreement was substantial (kappa = 0.74). When comparing machine versus AREDS, agreement was almost perfect (kappa = 0.85) and between the two graders it was substantial (kappa = 0.71).
This study demonstrates the feasibility of human-machine-based classification for automated AMD severity grading. Using one large public retinal image database, highly reproducible results were obtained when comparing manual and automated grading of severity of detected AMD. These techniques have the potential to provide individuals with automated grading of fundi to identify AMD or monitor those individuals with earlier stages of AMD for the onset of the more advanced stages when prompt treatment may be indicated to reduce the risk of blindness.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.
This PDF is available to Subscribers Only