International Journal of Imaging Systems and Technology Biomedical Engineering
Bayes DeepRetinaNet : A Breakthrough in Bayesian Analysis for Eye Health
Abstract
ABSTRACT Deep learning has significantly advanced automated retinal disease detection, yet standard convolutional neural networks (CNNs) often lack reliable uncertainty estimation, an essential property in high‐stakes clinical decision‐making. In this paper, we introduce Bayesian DeepRetinaNet, a probabilistic extension of DeepRetinaNet, incorporating Monte Carlo Dropout and Variational Inference for uncertainty‐aware prediction. We evaluate our models on two retinal fundus datasets: APTOS 2019 and DERBI Hackathon. Bayesian variants outperform the baseline CNN in terms of calibration, generalization, and robustness, particularly under noisy or ambiguous conditions. Moreover, qualitative uncertainty analysis reveals that Bayesian models express epistemic uncertainty more effectively, contributing to safer and more interpretable diagnostic outcomes. Our results highlight the clinical potential of integrating Bayesian learning into retinal image analysis pipelines.