Fig. 3.

Illustration of the progressive training-in-time approach. In the first level of training, the k-space data of all the frames are binned into one and we try to solve for the average image in this level. Upon the convergence of the first step, the parameters and latent variables are transferred as the initialization of the second step. In the second level of training, we divide the k-space data into M groups and try to reconstruct the M average images. Following the convergence, we can move to the final level of training, where the parameters obtained in the second step and the linear interpolation of the latent vectors in the second step are chosen as the initializations of the final step of training.