Abstract
Purpose :
Current deep learning (DL) models for age-related macular degeneration (AMD) primarily focus on identifying and quantifying features on optical coherence tomography. We developed a DL model to quantify intermediate AMD features on color fundus photographs (CFP). In this study, we evaluated the DL-based quantifications of intermediate AMD features in their ability to predict conversion to geographic atrophy (GA) and exudative AMD.
Methods :
In total, 37420 photographs of 17605 eyes of 8890 participants aged 45+ from the population-based Rotterdam Study were available for the prediction analysis. Drusen-, hyperpigmentation-, retinal pigment degeneration-, and reticular pseudodrusen areas on CFP were automatically segmented with the DL model. To predict conversion to GA and/or exudative AMD within a 7 year period, we used a random forest classifier and a generalized estimated equations (GEE) model adjusted for age, sex, subject and eye with 5-fold cross-validation. For comparison, we fitted similar models to human graded labels. Area under the operating receiving curves (AUCs) were constructed for the different late AMD subtypes.
Results :
173 eyes converted to GA and 172 to exudative AMD within a time period of 7 years. The random forest classifier based on automated intermediate AMD area quantifications reached a similar to higher AUC for conversion to GA (0.87 (95%CI 0.83-0.89) versus 0.85 [95%CI 0.82-0.89]) and for conversion to exudative AMD 0.79 (95%CI 0.75-0.83) versus 0.72 [95%CI 0.67-0.77]) when compared to human graded labels. The GEE showed similar results for GA (0.93 [95%CI 0.90-0.97] versus 0.94 [95%CI 0.90-0.98]) and for exudative AMD (0.87 [95%CI 0.85-0.90] versus 0.89 [95%CI 0.87-0.95]).
Conclusions :
DL-based quantifications of intermediate AMD features improve the prediction of conversion to GA and exudative AMD and can be used as a quick and objective method for prognostic modelling in patients at risk. Potential for use includes screening, epidemiologic studies, clinical management, and patient selection for clinical trials.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.