Learning a B₀ Shimming Model Using Deep Neural Networks

Magnetic resonance (MR) is a non-invasive imaging technique used in biomedical research and clinical care. Central to this imaging modality is a strong, homogeneous, and static magnetic field for achieving enhanced data quality and shorter acquisition times. However, various factors, such as tissue susceptibility differences and hardware imperfections, can introduce field inhomogeneities. B0 shimming is essential to compensate for field variations. We propose the use of deep neural networks (DNNs) to estimate shim coils coefficients, as DNNs can model complex patterns of field perturbation, learn implicit representation of the shim fields, and are typically very fast to evaluate at inference time. This paper presents a neural network-based shim model that generates optimal shim coefficients using a simulation dataset. The model can predict near optimal coefficients for arbitrary shim volume masks, achieving a performance of R²=0.941±0.005 in ideal and non-ideal shim conditions.Clinical RelevanceThe ability to quickly predict optimized shim coefficients at the scanner can potentially reduce scan time and enhance image quality in challenging imaging scenarios.