An untrained deep learning method for reconstructing dynamic MR images from accelerated model-based data.

The use of an untrained neural network together with a physics-based regularization loss shows promise as a measure for determining the optimal stopping point in training without relying on fully-sampled ground truth data.