Abstract
Purpose :
Prediction of age-related macular degeneration (AMD-)progression based on early imaging biomarkers is currently imprecise, but important for risk-based management and patient selection for clinical trials. Here, we evaluated the impact of drusen on AMD progression based on automated drusen area quantification.
Methods :
In total, 7964 participants aged 55+ from the population-based Rotterdam Study were followed up for AMD progression during a study period up to 15 years. Drusen areas within the early treatment diabetic retinopathy (ETDRS) grid on fundus photographs were automatically segmented with a deep learning algorithm at baseline, and were evaluated for their log correlation with consensus annotation by human graders. Associations between drusen area and incident late AMD were determined with cox proportional hazards models, adjusted for age and sex. Optimism-corrected area under the operating receiving curves (AUCs) were constructed for late AMD.
Results :
Drusen area quantifications by the deep learning model were highly correlated with consensus annotation by experienced graders (r = 0.81, p < 2.8e-10). During a mean follow-up of 6.7 years (SD 1.8), 107 participants developed incident late AMD. A 1 mm2 increase in deep learning-based drusen area in the ETRDRS grid was associated with a 4.9 times increased risk of incident late AMD (HR 4.9, 95%CI 3.3– 7.2). The prediction model based on automated drusen area quantification reached an AUC of 0.89 (95%CI 0.83-0.95).
Conclusions :
The deep learning model was on par with consensus annotation by human graders, and can be used as a quick and objective method for AMD end stage risk prediction. Automated quantification of early and intermediate AMD imaging biomarkers will enhance the investigation of causal relationships and patient selection for clinical trials.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.