Enhancing MRF Reconstruction: A Model-Based Deep Learning Approach Leveraging Learned Sparsity and Physics Priors

Deep learning has shown great promise in improving the speed and accuracy of parameter map estimation in magnetic resonance fingerprinting (MRF). However, many existing methods rely on physics-free networks, leading to a staged processing strategy. This strategy involves the initial reconstruction of acquired non-Cartesian undersampled measurements, followed by subsequent parameter map estimation. Unfortunately, such a staged processing strategy may lead to partial information loss and limit the eventual accuracy of parameter imaging. To overcome these challenges, in this paper, we propose a novel model-based deep learning approach that directly reconstructs parameter maps from non-Cartesian undersampled measurements. Specifically, our approach first incorporates MRF imaging physics priors and data correlation constraints into a unified reconstruction model. The proposed model-based network, named LS-MRF-Net, is then defined by unrolling the iterative procedures of the reconstruction model into a deep neural network. Notably, a learned sparsity layer is proposed to exploit the optimal transform domain for sparse representation of high-dimensional MRF data. Additionally, we incorporate a mapping layer and a Bloch response dynamic layer to seamlessly integrate the MRF imaging physics priors into the network. Experimental results on both simulated and in vivo datasets demonstrate that the proposed method can significantly reduce computational time while enhancing MRF reconstruction performance.