Dictionary learning for compressive parameter mapping in magnetic resonance imaging

Parameter mapping is a valuable quantitative tool for soft tissue contrast. Accelerated data acquisition is critical for clinical utility, which has lead to various novel reconstruction techniques. In this work, a model-based compressed sensing method is extended to include a sparse regularization that is learned from the principal component coefficient. The principal components for a range of T² decay curves are computed, and the coefficients of the principal components are reconstructed. These coefficient maps share coherent spatial structures, suggesting a patch{based dictionary is a well suited sparse transformation. This transformation is learned from the coefficients themselves. The proposed reconstruction is suited for non-Cartesian, multi-channel data. The dictionary constraint leads to parameter maps with less noise and less aliasing for high amounts of acceleration.