Investigating the Efficacy of Nonlinear Susceptibility Inversion Deep Learning Models in MRI Quantitative Susceptibility Mapping

Abstract:
Quantitative Susceptibility Mapping (QSM) serves as a crucial tool for assessing tissue susceptibility in various cerebral conditions. Derived from MRI gradient-echo signals, QSM delineates tissue susceptibility distribution by solving a complex dipole inversion problem. However, traditional QSM reconstruction methods often yield images marred by streaking artifacts, excessive smoothing, noise amplification, or underestimated susceptibility values. Moreover, datadriven deep learning-based QSM approaches often lack a data fidelity component within their network architecture, leading to deviations from dipole model-derived magnetic susceptibility values on original phase images. To mitigate these challenges, this paper introduces a novel nonlinear model-driven deep learning method termed NDL-QSM. NDL-QSM incorporates a nonlinear susceptibility inversion model into a deep learning framework and employs the PGD (Projected Gradient Descent) algorithm for systematic optimization. Leveraging the merits of nonlinear modeling and deep learning networks, NDL-QSM delivers more precise susceptibility estimations and produces higher-quality susceptibility maps compared to existing methodologies.