Joint Optimization of Sampling Pattern and Reconstruction for Dynamic MRI

Deep learning (DL) methods have shown promising results at solving accelerated dynamic magnetic resonance imaging (MRI) reconstruction problems. However, the sampling patterns used for in DL reconstruction studies are typically chosen heuristically. The reconstruction models are optimized for a pre-determined acquisition (encoding) model without taking advantage of the interaction between data sampling and reconstruction. In order to capture the spatio-temporal characteristics more effectively, we jointly optimize <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$k-t$</tex> sampling patterns and reconstruction networks by extending the recent AutoSamp framework to dynamic MRI setting. Experiments on retrospectively and prospectively undersampled cardiac cine data show that our method outperforms traditional heuristic sampling approaches, consistently improving image quality across various acceleration factors. Overall, our deep-learning based approach improves the efficiency of dynamic MRI acquisition and reconstruction, accurately capturing complex spatio-temporal dynamics.