Motivation: The prior knowledge of the latent manifold structure implicit in the imaging principles can be further explored for improved MRF reconstruction. Goal(s): To propose a novel deep-learning framework with manifold structure priors of MRF data and other data priors for improved MRF reconstruction. Approach: We propose a novel deep unrolled network based on manifold structured data regularization in the non-Euclidean norm sense. In addition, we impose additional sparsity constraints on the parameter maps to further improve the accuracy of the manifold structure estimation. Results: Experimental results demonstrate that the proposed method outperforms the original manifold structured data priors-based method and several state-of-the-art methods. Impact: By further incorporating the manifold structure priors along with the data priors in the parameter domain, our method can provide more accurate tissue quantification.
Ji et al. (Tue,) studied this question.