Scholastic Bias-Accelerated Subset Selection Algorithm for Joint Learning of Sampling Pattern And Reconstruction In Accelerated MRI

This work proposes and illustrates a stochastic version of the bias-accelerated subset selection (BASS) algorithm used to learn the sampling pattern (SP) in accelerated Cartesian MRI problems. Recently, BASS was combined with ADAM for jointly learning of the SP and the deep learning reconstruction. However, BASS originally uses all the data in each iteration, taking a long processing time when the training dataset is large. Here, we illustrate that BASS is also very stable when only a small fraction of the dataset is used at each iteration. This version, called stochastic BASS (SBASS), can substantially reduce the SP training time for large datasets, obtaining similar quality measurements of the learned pair of SP and reconstruction.