Uncertainty-aware diabetic retinopathy detection using deep learning enhanced by Bayesian approaches

TL;DR

A study that applied Bayesian approximation techniques (MC Dropout, Mean Field Variational Inference, Deterministic Inference) to DenseNet-121 for diabetic retinopathy detection, improving both prediction accuracy and uncertainty estimation.

Problem

Deep learning models in medical image analysis show high performance but fail to provide confidence information in uncertain cases due to single-point predictions. In diabetic retinopathy detection, even small errors can have serious impacts, making uncertainty quantification crucial.

Approach

A pre-trained DenseNet-121 was fine-tuned on diabetic retinopathy data (APTOS 2019 + DDR) via transfer learning, then Bayesian approximation methods (MC Dropout, Mean Field Variational Inference, Deterministic Inference) were applied to approximate the posterior predictive distribution and estimate uncertainty. Uncertainty was measured using entropy and standard deviation, and AUC and accuracy were evaluated at multiple data retention levels.

Results & Contribution

The MC Dropout model achieved the highest test accuracy of 97.68%, followed by Mean Field Variational Inference (94.23%) and the Deterministic model (91.44%). The Bayesian approach not only improved classification accuracy but also provided uncertainty information, aiding clinical decision-making.