Data-driven regularization parameter selection in dynamic MRI
In dynamic MRI, sufficient time resolution can often only be obtained using imaging protocols which produce undersampled data for each image in the time series. This has led to the popularity of compressed sensing (CS) based image reconstruction approaches. One of the problems in CS approaches is determining the regularization parameters, which control the balance between data fidelity the spatial and temporal regularization terms. A data-driven approach is proposed for the total variation regularization parameter selection such that the reconstructions yield expected sparsity levels in the regularization domains. The expected sparsity levels are obtained from the measurement data for the temporal regularization and from a reference image for the spatial regularization. Two formulations are proposed. The first is a 2D search for a parameter pair which produces expected sparsity in both the temporal and spatial regularization domains. In the second approach, the sparsity-based parameter selection is split to two 1D searches using the S-curve method. The approaches are evaluated using simulated and experimental DCE-MRI. In the simulated test case, both proposed methods produce a parameter pair that is close to the RMSE optimal pair, and the reconstruction error is also close to minimum. In the experimental test case, the methods produce almost similar parameter selection, and the reconstructions are of high perceived quality. Both approaches lead to a highly feasible selection of the temporal and spatial regularization parameters in both the simulated and experimental test cases while the sequential method is computationally more efficient.