Physics-Driven Learned Deconvolution of Multi-Spectral Fluorine-19 MRI with Multiple Agents Using Radial Sampling

Motivation: Detection of multiple cell targets separately labeled with different chemically-shifted, paramagnetic 19F tracers can benefit from radial k-space sampling pulse sequences; however, radial sampling can lead to non-linear smearing chemical shift artifacts. Goal(s): Our goal is to develop suitable modeling of the radial chemical shifts and a physics-informed deconvolution scheme to unmix multi-spectral components. Approach: We proposed a novel Radon transform modeling of forward operator for radial chemical shifts and introduced machine learning based method for multi-spectral deconvolution. Results: Radial chemically-shifted artifacts are significantly reduced via the deep unrolling learned deconvolution algorithm, especially for low signal-to-noise-ratio (SNR) and highly undersampled acquisitions. Impact: Unlike Cartesian chemical shifts that result in image displacements, radial chemical shifts produce more complex artifacts. To effectively unmix the multi-spectral 19F components, we developed an analytical model using Radon transform and a data-driven deconvolution method based on deep unrolling.