Uncertainty Analysis of Deep Kernel Learning Methods on Diabetic Retinopathy Grading

Diabetic retinopathy (DR) is a leading cause of vision loss. Therefore, screening for early signs and assessment of the DR severity is crucial and extensively studied. To support clinicians, screening could be automated by algorithms that can, for example, refer difficult decisions to specialists for further investigation. However, frequently used neural networks (NNs) typically do not know when they do not know and approximate Bayesian NNs often equally do not suffice to provide well-calibrated uncertainty estimates. Thus, in this work, we investigate whether and how to use deep kernel learning (DKL) which we designed as a hybrid combination of the state-of-the-art EfficientNet-B0 and a Gaussian process (GP) layer to improve the quality of uncertainty estimates in referral-based DR screening. To this end, we first analyze the necessity for recently proposed extensions to the DKL framework to resolve miscalibrated uncertainties, despite the theoretical superiority of GPs to uncertainty quantification. Our subsequent comprehensive comparison of the curated DKL's performance to that of the most common approximate Bayesian NNs shows our DKL models to particularly improve the detection of near out-of-distribution (OOD) samples containing other eye diseases through epistemic uncertainty information, but also enhance the calibration of aleatoric in-distribution uncertainty and diagnostic performance. Hence, it can provide a substantial benefit for safety-critical medical applications, like automated DR screening, particularly by potentially reducing the risk of missing diseases other than DR due to the improved near-OOD detection performance.