Figure 1.

(A) A generalized image reconstruction algorithm iteratively removes aliasing artifacts produced by undersampling. The methods alternate between data consistency and regularization steps. A common conventional approach (e.g., [22]) would apply regularization via a proximal step. DL-based reconstruction methods regularize via a neural network. (B) An alternative reconstruction algorithm uses the partial separability model. Image reconstruction is performed on spatial and temporal basis rather than directly on the images. A conventional method (e.g., [47,48]) alternately updates the basis U and V via gradient steps. Similarly, a DL-based reconstruction method would update via spatial and temporal networks. Diagrams and pictures modified from Sandino et al. [49].