FIGURE 15.

An illustration of the SPICE approach. Each voxel spectra in the high‐dimensional spatiospectral function of interest (ρ(x,f), the image on the left) is modeled as a linear combination of a small number of spectral basis functions $({\left\{{\phi }_l\left(f\right)\right\}}_{l=1}^L$, rightmost column). This implies that the high‐dimensional signals reside in a low‐dimensional subspace (spanned by {φ_l(f)}). With the subspace predetermined (from training data), the imaging problem is transformed into the estimation of a set of spatial coefficients ( ${\left\{c_l\left(x\right)\right\}}_{l=1}^L$) with much lower dimensions than the original spatiospectral function. Different subspaces can be constructed for different signal components, i.e., water, lipids and metabolites