Purchase this article with an account.
Arno P Goebel, Stefan Saur, Christian Wojek, Christoph Russmann, Frank G Holz, Steffen Schmitz-Valckenberg, MODIAMD study group; Automated retinal image analysis for evaluation of high-risk characteristics in intermediate age-related macular degeneration. Invest. Ophthalmol. Vis. Sci. 2014;55(13):5206.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To investigate a newly-developed software tool for retinal imaging analysis of high-risk characteristics in patients with intermediate age-related macular degeneration (AMD).
Multi-modal retinal imaging including color fundus photography (CFP) and fundus autofluorescene (FAF) was obtained at baseline (t0) and at one year later (t1) in 55 eyes of 48 patients with AMD (median age 73). Automated image registration was performed for the two CFP images (t0 and t1) and for the two pairs of CFP and FAF images (for t0 and t1, respectively). Registration accuracy was measured by pixel distance errors of point pairs manually set by an expert reader in all the images based on anatomical landmarks. For high-risk feature identification, an algorithm has been developed for automated detection of hyperpigmentary changes. A statistical model was trained using (a) only image features from the CFP images and (b) multimodal image features, thereafter tested on CFP images not used for training.
Accuracy of the registration algorithm was superior (p<0.01) to manual registration. The median [95% CI] error between CFP images (t0 to t1) was 3.7 pixels [0.7, 10.6] as compared to 4.8 pixels [0.9, 15.9] for manual registration. For automated FAF to CFP registration, the median error was 4.6 pixels [0.9, 20.0] for t0 and 4.8 pixels [0.9, 15.4] for t1 as compared to 5.1 pixels [0.9, 16.4] for t0 and 5.3 pixels [01.1, 19.6] for t1 after manual registration. Automated registration required appr. 10 sec (CFP) or 30 sec (FAF to CFP) in a research prototype, whereas it took a human expert appr. 45 sec for the manual registration of two images (CFP or FAF to CFP). The sensitivity at a specificity level of 98% and 99% for hyperpigmentation detection (manual annotation as reference) by the research prototype on a pixel level was 92.9% and 86.7% using only CFP images for training (a) and 93.3% and 87.2% for using both CFP and FAF images (b) with significant (p=0.01) differences between both approaches.
By means of automated registration multimodal and longitudinal characteristics of high-risk AMD features can readily be recorded in a non time-consuming manner. The research prototype allows for automated detection of focal hyperpigmentary changes, which will be useful both in future natural history and interventional studies.
This PDF is available to Subscribers Only