BAY-MED: Bayesian Approximation for Post-Hoc Uncertainty in Medical Imaging

Uncertainty quantification (UQ) is crucial in deep learning for medical image analysis, as reliable uncertainty estimates enhance trust and support decision-making. Traditional UQ methods, such as Bayesian Neural Networks and Monte Carlo Dropout, often require model modifications or multiple forward passes, increasing computational complexity. We propose a post-hoc UQ method that employs a meta-model trained on features from a pre-trained model, offering computational efficiency without altering the original architecture. Our approach leverages an Dirichlet evidential lower bound (D-ELBO) loss to effectively capture both aleatoric and epistemic uncertainties. We trained our base and meta-models on the BACH dataset (indistribution) and evaluated their performance using three out-of-distribution (OOD) imaging datasets-DIV2K, HAMIOOOO, and BreakHIS-to assess OOD detection and uncertainty estimation. Our method demonstrates superior performance in calibration and OOD detection, achieving improved mutual information and precision scores. This computationally efficient UQ method can be seamlessly integrated with existing models to enhance the reliability, safety, trustworthiness, and interpretability of medical AI systems in clinical applications.