Unsupervised Fixed Point Networks for Improved MRF Reconstruction

Magnetic resonance fingerprinting (MRF) is a powerful quantitative imaging technique that simultaneously estimates multiple tissue parameters from a single acquisition. However, severe undersampling in MRF often introduces artifacts and noise, compromising reconstruction accuracy and reliability. To address this, we propose UFP-MRF, a novel unsupervised fixed-point network for MRF reconstruction. The framework integrates the physical acquisition model with learnable regularization components, ensuring strong convergence through contractive mapping design. Leveraging a deep equilibrium (DEQ) formulation, UFP-MRF efficiently computes fixedpoint solutions that implicitly capture infinite-depth network behavior, achieving high-quality reconstructions at equilibrium. Trained via a data-consistency loss without fully sampled ground truth, UFP-MRF rapidly converges to a stable solution and consistently outperforms state-of-the-art methods in both reconstruction accuracy and robustness.