Uncertainty-Aware Geographic Atrophy Progression Prediction from Fundus Autofluorescence

Geographic atrophy (GA) is an advanced form of age-related macular degeneration leading to progressive visual loss. The ability to accurately predict GA progression over time based on a single baseline visit can improve clinical trials in GA, as well as support patient counseling in current clinical practice. The feasibility of using baseline fundus autofluorescence (FAF) images to predict GA progression with end-to-end deep learning models has been demonstrated. However, for black-box models, there is a need to increase trust for clinical practice applications and estimate the prediction uncertainty. In this paper, we applied and evaluated both non-parametric and parametric deep ensemble approaches for the prediction uncertainty estimation using both simulated and clinical study data in a multitask regression setting. The results not only show promising performance in detecting near and far out-of-distribution data cases, but may also suggest the improved performance in predicting GA growth rate for in-distribution data.