A unified approach for maintaining MRI reconstruction quality and quantifying both aleatoric and epistemic...

This study investigates practical design choices for Bayesian uncertainty quantification (UQ) in model-based deep learning (MoDL) for accelerated MRI reconstruction. Specifically, we examine where to introduce stochastic layers to capture epistemic uncertainty while maintaining reconstruction quality, and how to integrate heteroscedastic learning to capture aleatoric uncertainty. We compare Monte Carlo dropout (MCD) and Bayes-by-Backprop (BBB) across multiple layer placement patterns and propose a two-stage training strategy in which the reconstruction network is frozen before training the aleatoric uncertainty network. Experiments demonstrate that the placement of stochastic layers significantly impacts performance for MCD, whereas BBB exhibits greater robustness. Furthermore, the proposed two-stage strategy yields statistically significant improvements in peak signal-to-noise ratio compared with conventional joint training. Under representative domain shifts (anatomical and acceleration-factor shifts), the estimated uncertainties capture expected in-domain versus out-of-distribution trends, validating the proposed framework. These results provide implementation-level guidance for realizing effective Bayesian UQ in MoDL-based MRI reconstruction.