UMPIRE-Net decouples magnitude and phase for sharper accelerated MRI

MRI reconstruction from undersampled k-space is an ill-posed inverse problem, where physics-driven deep learning (PD-DL) methods combine the MRI forward model with learned regularization inside algorithm-unrolling frameworks. However, most PD-DL approaches reconstruct complex-valued images directly, implicitly coupling magnitude and phase in a single representation. This is suboptimal for settings like partial Fourier (PF) imaging, where accurate phase modeling is critical for recovering omitted asymmetric k-space measurements. UMPIRE-Net (Unrolled Magnitude-Phase In REgularization Network) tackles this by introducing separate learned regularizers for the magnitude and phase components, along with a novel data-fidelity term that enforces measurement consistency. This explicit decoupling reduces reliance on externally estimated or predefined phase information.

Evaluated on accelerated MRI with PF across multiple datasets and acceleration factors, UMPIRE-Net consistently outperforms conventional complex-valued PD-DL baselines. The results show sharper images and reduced artifacts, demonstrating that explicit modeling of magnitude and phase improves reconstruction quality in scenarios where phase accuracy matters. The code is publicly available on GitHub, making it easy for other researchers to reproduce and build upon. The paper was accepted at the IEEE International Workshop on Machine Learning for Signal Processing (MLSP), underlining its relevance to the signal processing community.