Rapid T₁ quantification from high resolution 3D data with model‐based reconstruction

Abstract

Purpose Magnetic resonance imaging protocols for the assessment of quantitative information suffer from long acquisition times since multiple measurements in a parametric dimension are required. To facilitate the clinical applicability, accelerating the acquisition is of high importance. To this end, we propose a model‐based optimization framework in conjunction with undersampling 3D radial stack‐of‐stars data. Theory and Methods High resolution 3D T 1 maps are generated from subsampled data by employing model‐based reconstruction combined with a regularization functional, coupling information from the spatial and parametric dimension, to exploit redundancies in the acquired parameter encodings and across parameter maps. To cope with the resulting non‐linear, non‐differentiable optimization problem, we propose a solution strategy based on the iteratively regularized Gauss‐Newton method. The importance of 3D‐spectral regularization is demonstrated by a comparison to 2D‐spectral regularized results. The algorithm is validated for the variable flip angle (VFA) and inversion recovery Look‐Locker (IRLL) method on numerical simulated data, MRI phantoms, and in vivo data. Results Evaluation of the proposed method using numerical simulations and phantom scans shows excellent quantitative agreement and image quality. T 1 maps from accelerated 3D in vivo measurements, e.g. 1.8 s/slice with the VFA method, are in high accordance with fully sampled reference reconstructions. Conclusions The proposed algorithm is able to recover T 1 maps with an isotropic resolution of 1 mm 3 from highly undersampled radial data by exploiting structural similarities in the imaging volume and across parameter maps.

Spatially regularized estimation of the tissue homogeneity model parameters in DCE‐MRI using proximal minimization

“…This case can be computed again by a proximal method,such as the nested primaldual algorithm 27 described in Section 2.4.2. Let us note that in the case of a nonconvex d, as in our case, the proximal Newton method converges to the closest minimum, 44,45 if the minimization steps are not too large.…”

Improvement of Fast Model-Based Acceleration of Parameter Look-Locker T1 Mapping

“…In order to get better time of computational we decided to choose a simple and efficient method of model fitting. Other solutions, in some cases are more efficient but are also numerically advanced, and require more time to reconstruct the final image [16][17][18][19][27][28][29]33]. In the literature we did not find any benchmark that enables us to compare methods of T 1 mapping reconstruction and quality of results of different works for similar data.…”

“…Acceleration using GPU and implementation of some methods in C/CUDA (Compute Unified Device Architecture) generates a time complexity of 10-20 min [19,33] and in the range from minutes to hours depending on data size [18]. Fast multi-slice method for T 2 mapping [29] calculates 50 slices on an office computer within 7 h, which gives approximately 9 min per slice or alternatively rapid T 1 quantification [28] reports reconstruction time of approximately 10 min per slice. We showed that a single iteration of the FIR-MAP for all projections could take 30 s. The data preparation of initial model for all projections took an additional 2 min for IFM and 1 min for MFM.…”

“…Zibetti et al provides comparison and evaluation of 12 different types of CS sparsity for acceleration of T 1 mapping [27]. More recent methods use improvements of previous algorithms by means of total-generalized-variations (TGV) based regularization and further adapted to a multiparametric setting [28] or split-slice GRAPPA and a model-based iterative algorithm for T 2 -mapping [29]. Different approach bases on the method of magnetic resonance fingerprinting [30,31] in which the benefit comes from simultaneous computation of T 1 and T 2 maps but in slightly longer acquisition time.…”

“…Most of the current works operates on a single slice, which from a clinical point of view is not applicable. The limited methods of multi-slice parameter mapping have been used in few works [28,29,33] resulting in calculation time from 10 min [28] up to 7 h [29] in multi-slice analysis. In multi-channel systems the number of coils may be limited in the preprocessing step.…”