Meeting Banner
Abstract #2538

Automatic Regularization Parameter Selection for Iterative Nonlinear MRI Reconstruction

Sathish Ramani1, Jon-Fredrik Nielsen2, Jeffrey A. Fessler1

1EECS, University of Michigan, Ann Arbor, MI, United States; 2fMRI Laboratory, University of Michigan, Ann Arbor, MI, United States

MRI reconstruction from undersampled sampled k-space data requires regularization to reduce artifacts. Nonquadratic regularization (e.g., l1 or TV) can be used to restore image quality, but its successful application depends on proper selection of the regularization parameter. In this work, we demonstrate the applicability of generalized cross validation (GCV) and an estimate of a mean-squared error (MSE)-type measure for quantitative selection of the regularization parameter for MRI reconstruction using the iterative split-Bregman algorithm with nonquadratic regularization. GCV and the MSE-type estimate require the Jacobian matrix of the reconstruction with respect to the data that we evaluate analytically for the SB algorithm. We illustrate with experiments on real MR (phantom) data that GCV and the MSE-type estimate lead to near-optimal reconstruction results.

Abstract PDF Poster PDF

Keywords

able adjustment analytical analytically application arbor artifacts assuming automated automatic available away become chain close complex components computation compute computing conjugate constant constraint controls corresponds cost criteria cross dashed denotes dependence depends desired dummy employ employed encode enormous error estimation evaluate evaluated exactly exhibits experiment explicit feasibility finite form fully functions generalized governs gradients handle health illustrated indicate instances involving iteratively laboratory linear matrices matrix measures minimization minimize minimizing national near nonlinear norm object operator optimal original parallel part phantom plot popular practical practice predicted principle prob problem products proper propose quality random real reconstruct reconstruction reconstructions recursively reduce reduced regime regularization relatively represents require requires respectively retrospective retrospectively risk rule sample sampled scale selection sequel served sets setting several simple smoothness soft space split splitting squared stein stochastically store sure surf thresholding trace transform transformation transpose treating true tuned type typical update variable variance variation vector vectors versus written zero