ADEPT: Accurate Diffusion Echo‐Planar imaging with multi‐contrast shoTs

Abstract

Theory and Methods:We propose a new framework called ADEPT (Accurate Diffusion Echo-Planar imaging with multi-contrast shoTs) that enables fast diffusion MRI by allowing diffusion contrast settings to change between shots in a multi-shot EPI acquisition (i.e., intra-scan modulation). The framework estimates diffusion parameter maps directly from the acquired intra-scan modulated k-space data, while simultaneously accounting for shot-to-shot phase inconsistencies. The performance of the estimation framework is evaluated using Monte Carlo simulation studies and in-vivo experiments and compared to that of reference methods that rely on parallel imaging for shot-to-shot phase correction.Results: Simulation and real-data experiments show that ADEPT yields more accurate and more precise estimates of the diffusion metrics in multi-shot EPI data in comparison with the reference methods. Conclusion: ADEPT allows fast multi-shot EPI diffusion MRI without significantly degrading the accuracy and precision of the estimated diffusion maps.

Knowledge‐driven deep learning for fast MR imaging: Undersampled MR image reconstruction from supervised to un‐supervised learning

“…[125][126][127][128][129][130][131][132][133][134][135][136][137][138][139] Developing supervised DL reconstruction methods for qMRI is more challenging than for conventional MRI since fully sampling multiple contrast weighted images would lead to infeasible scan time. 139 However, the physical model of qMRI and its flexible representations such as analytical equations, dictionary, linear subspace, or even the original Bloch equation provides the opportunities for designing self-supervised or unsupervised DL reconstruction methods. There has been a review article pointing out the typical four ways to integrate DL with the quantitative MRI.…”

Systematic review of reconstruction techniques for accelerated quantitative MRI

“…Developing novel models that parameterize the model jointly for auxiliary parameters. For instance: the joint estimation of off‐resonance frequency in $$$ {\mathrm{T}}_2^{\ast } $$$ mapping, 279 the joint estimation of $$$ {\mathrm{T}}_2^{\ast } $$$ and field maps in water and fat parameter estimation, 272,280,281 the joint estimation of flip angles and proton density maps in $$$ {\mathrm{T}}_2 $$$ mapping, 282 the joint estimation of steady‐state signal, equilibrium signal, and effective relaxation rate in $$$ {\mathrm{T}}_1 $$$ mapping, 283‐285 the joint estimation of tracer kinetic model parameters in $$$ {\mathrm{T}}_1 $$$ mapping, 286 and the joint estimation of phase parameters in multishot diffusion imaging 287,288 …”

“…For instance: the joint estimation of off-resonance frequency in T * 2 mapping, 279 the joint estimation of T * 2 and field maps in water and fat parameter estimation, 272,280,281 the joint estimation of flip angles and proton density maps in T 2 mapping, 282 the joint estimation of steady-state signal, equilibrium signal, and effective relaxation rate in T 1 mapping, [283][284][285] the joint estimation of tracer kinetic model parameters in T 1 mapping, 286 and the joint estimation of phase parameters in multishot diffusion imaging. 287,288 Regularized direct reconstruction. To improve the performance of direct reconstruction, prior information about the qMRI parameters to be estimated can be incorporated by adding regularization terms, such as Tikhonov regularizer, 274 l 1 norm regularizer, 289 and weighted l 1 ball regularizer 268,290 to the cost function described by Equation (24).…”

“…278,309 The computational time of techniques in direct reconstruction can be decreased using block-wise decomposition of the cost function in combination with a specific uniform undersampling pattern. 310 In case multiparametric models are used, advanced solvers such as Stochastic Gradient Langevin dynamics, 286 regularized trust region continuation techniques, 279 or alternating minimization 288,295 have been proposed to optimize the nonlinear cost-function jointly for different parameters. Furthermore, it has been proposed to automatically scale the unknown parameters in the joint optimization, which numerically balances the partial derivatives of the multiparametric model to improve the conditioning and hence ease the optimization.…”