Speaker
Description
State space models (SSMs) have recently emerged as efficient alternatives to Transformers for MRI reconstruction, offering linear complexity while maintaining global receptive fields. LMO (Linear Mamba Operator, Li et al., CVPR 2025) formulates MRI reconstruction as operator learning in bandlimited function spaces, achieving competitive results on standard benchmarks. We identify a physical modeling error in LMO's data consistency (DC) mechanism: the DC layer operates on magnitude images, discarding phase information before enforcing k-space consistency with complex-valued measured data. This mixes incompatible frequency-domain representations at every DC step, introducing systematic errors throughout the network. Notably, this behavior is undocumented — it is only visible in the released code. We propose a physically grounded correction that preserves the SSM architecture: (1) sensitivity-weighted complex input (2-channel real/imaginary), (2) a ComplexDC layer applying proper complex-valued k-space enforcement, and (3) adapted lifting and projection layers — adding only 1.4% parameters. On FastMRI Brain (Poisson 2D, 8x acceleration), the corrected model yields +1.74 dB PSNR, +0.012 SSIM, and -24% NMSE over the magnitude-domain baseline, with only 2.15M parameters — less than half of the original LMO (~5M) and a fraction of UNet (31M). Our results suggest that physical correctness of data consistency is a more effective lever for reconstruction quality than architectural refinements to the SSM itself, particularly in data-limited regimes.