Benchmarking Physics-Informed Unrolled Reconstruction for Accelerated Low-Field Brain MRI on a 47 mT Open-Source System

Low-field MRI widens imaging access in resource-constrained settings, but low signal-to-noise ratio and long scans motivate k-space undersampling, requiring reconstruction that is accurate and computationally light. We present a physics-informed unrolled network for accelerated reconstruction on a 47 mT open-source scanner (OSI2 ONE). It interleaves a shared residual regularizer with conjugate-gradient data-consistency layers using only 0.37 M parameters. We evaluate conservatively: retrospective undersampling, brain-region metrics, subject-level statistics, leave-one-subject-out cross-validation, and multiple seeds. At fourfold acceleration the model improves SSIM by 0.097 and PSNR by 3.0 dB over zero-filling (subject-level paired test, p < 0.01) and exceeds total-variation and learned baselines (U-Net, ISTA-Net) at every rate, while matching a shared-weight variant that isolates the role of data consistency. Accuracy saturates at four cascades, and a volume reconstructs in under one second on a GPU using 33 MB of memory. On genuinely acquired accelerated scans the method ties zero-filling against a noisy reference, as do all baselines, a tie we attribute to the limits of full-reference metrics rather than reconstruction quality.