Nonlinear inversion model-driven deep learning method for magnetic resonance imaging (MRI) quantitative susceptibility mapping imaging

Background: Quantitative susceptibility mapping (QSM) has great advantages in evaluating tissue susceptibility across diverse cerebral conditions. However, conventional reconstruction methods are often affected by streaking artifacts and noise amplification, while purely data-driven deep learning approaches frequently lack physical constraints, resulting in deviations from the underlying dipole physics. To address these issues, we aimed to develop a model-driven deep learning approach that explicitly enforces dipole model data fidelity within the network, aiming to enhance quantitative accuracy and suppress artifacts. Methods: We propose a nonlinear susceptibility inversion deep learning model (NSIDL), which integrates a nonlinear susceptibility inversion (NSI) model into a convolutional neural network and employs the proximal gradient descent (PGD) method to solve the optimization problem. The method was trained and validated using a multi-orientation gradient echo magnetic resonance imaging (MRI) dataset. Quantitative performance of NSIDL was evaluated using the reconstruction challenge (RC-1 and RC-2) datasets and in vivo data, and compared with state-of-the-art iterative methods and deep learning approaches. Clinical feasibility was assessed in patients with hemorrhage, calcification, and multiple sclerosis (MS). Results: range, 0.3677-0.5714). On the RC-1 dataset, NSIDL demonstrated superior image fidelity, exhibiting the lowest normalized root mean square error (NRMSE) (63.267±0.575) and high-frequency error norm (HFEN) (58.300±0.668), and the highest peak signal-to-noise ratio (PSNR) (42.39±0.311), consistently outperforming other methods across all three metrics (P