Accelerated Iterative Reconstruction of Temporally Regularized Dynamic MRI

Introduction: The fundamental challenge in dynamic MR imaging is the tradeoff between spatial resolution and temporal resolution. Most traditional dynamic image acquisition methods and associated image reconstruction methods have been based on k-space operations. One acquires a temporal sequence of incomplete k-space sample sets, along with one (or two) complete reference datasets, and then employs some form of data sharing between the sets to find the “missing” data. Finally an inverse FFT is applied to the imputed k-space datasets to form the dynamic image sequence. The Keyhole method and reduced-encoding imaging with generalized-series reconstruction (RIGR) are two examples of data sharing techniques [1,2]. Both Keyhole and RIGR are based on the implicit assumption that the object varies smoothly over time. In this work, we use a modelbased reconstruction method that does not require recovering data in the frequency domain, but rather relies on explicit temporal interpolation in the image domain. Similar formulations have been studied in electrocardiography [3-5], and applied to simulated cardiac MRI data [6]. This method requires minimization via the iterative Conjugate Gradient (CG) algorithm, which requires more computation than traditional keyhole methods. To reduce computation, we present an accelerated CG algorithm for dynamic imaging based on Toeplitz matrices and FFT operations [7].