Uncertainty-weighted dual-domain consistency for self-supervised accelerated MRI reconstruction

Abstract

Accelerated magnetic resonance imaging (MRI) typically relies on fully sampled k-space data to train supervised reconstruction networks, which are difficult to acquire in clinical practice. To eliminate this dependency, we propose a novel dual-domain self-supervised learning (Dual-SSL) framework that rigorously enforces physical consistency in both image and k-space domains. The framework employs a shared-intersection re-undersampling strategy to generate two sub-masks with approximately 50% overlap, preserving essential low-frequency priors, which are subsequently concatenated for efficient single-forward-pass processing. A composite loss function is formulated to enforce image-domain consistency alongside comprehensive k-space consistency, the latter encompassing cross-consistency on mutually unobserved points and self-consistency on originally observed points. To automatically balance these dual-domain constraints without empirical hyperparameter tuning, a homoscedastic uncertainty-based mechanism is incorporated to learn log-variance parameters. The proposed framework was evaluated on fastMRI and MICCAI 2013 datasets under 1D Cartesian and 2D random variable-density undersampling at 4× and 8× accelerations. Results demonstrate that this approach significantly outperforms existing self-supervised techniques in mitigating severe aliasing artifacts and localized error hotspots, thereby yielding superior anatomical fidelity. Crucially, despite training exclusively on undersampled data, the proposed framework achieves reconstruction performance comparable to fully supervised baselines using the identical network architecture, effectively bridging the gap between self-supervised and supervised MRI reconstruction. This provides a robust and scalable solution for clinical MRI acceleration.