Bayesian deep learning applied to diabetic retinopathy with uncertainty quantification

Deep Learning (DL) has significantly contributed to the field of medical imaging in recent years, leading to advancements in disease diagnosis and treatment. In the case of Diabetic Retinopathy (DR), DL models have shown high efficacy in tasks such as classification, segmentation, detection, and prediction. However, DL model's opacity and complexity lead to errors in decision-making, particularly in complex cases, making it necessary to estimate the model's uncertainty in predictions. Therefore, there is a need to estimate uncertainty in the model's predictions, which cannot be estimated by classical DL models alone. To address this issue, Bayesian DL methods have been proposed, and their use is increasing in the field. In this paper, we developed a straightforward architecture for the classification of DR using a Convolutional Neural Network (CNN) model. We then applied the Bayesian CNN twice, once using Variational Inference (VI) and once using Monte Carlo dropout (MC-dropout) methods, to the same CNN architecture. This allowed us to gain the posterior predictive distributions for each of them. The performance of the proposed models was evaluated on two benchmark datasets, namely APTOS 2019 and Messidor-2. Experimental findings demonstrated that the proposed models surpassed other state-of-the-art models, achieving a test accuracy of 94.70 % and 77.00 % for CNN, 94.00 % and 86.00 % for BCNN-VI, and 93.30 % and 81.00 % for BCNN-MC-dropout on the APTOS dataset and Messidor-2 dataset, respectively. Finally, we computed the entropy and standard deviation on the obtained predictive distribution to quantify the model uncertainty. This research highlights the potential benefits of using Bayesian DL methods in medical image analysis to improve the accuracy and reliability of diagnosing disease and treatment.