Motivation: Signal dynamics in MRI are influenced by multiple factors at different scales, complicating analyses. Existing analytical solutions and simulation methods are inadequate for complex geometries or susceptibility distributions. Goal(s): We aim to demonstrate and validate a physics-informed neural network (PINN) capable of handling arbitrary susceptibility-induced Larmor fields and diffusion terms. Approach: We implemented a flexible PINN-based simulation framework for solving the underlying the Bloch-Torrey equation which is able to predict the signal dynamics over time. Results: The framework produced accurate results for local and total magnetization across various Larmor field configurations at low to moderate oscillation frequencies, closely matching finite difference method results. Impact: By accommodating complex susceptibility distributions and geometries, our PINN-based framework enhances the simulation of MRI signal dynamics in tissues with varying properties —such as hemorrhages, amyloid deposits, iron accumulation, and vascular malformations —offering significant clinical relevance.
Rotkopf et al. (Tue,) studied this question.