Reconstruction of High-Resolution Metabolite Maps from Noisy MRSI Data by Incorporating Spatiospectral Constraints through Learned Kernels

High-resolution MR spectroscopic imaging (MRSI) suffers from very low signal-to-noise ratio, which is often addressed using a priori information/constraints. Existing constrained reconstruction methods utilize spectral constraints in the form of spectral subspaces/manifolds, while impose spatial constraints though spatial regularization. This paper presents a novel kernel-based partial separability model for reconstruction of high-resolution of metabolite maps from noisy MRSI data. The proposed model uses spectral basis functions to absorb spectral prior and a learned kernel function to absorb spatial prior. Experimental results demonstrated very encouraging reconstruction performance.