Highly Accelerated Chemical Exchange Saturation Transfer (CEST) Imaging by Combining Parallel Imaging and Compressed Sensing at 3T


Chemical Exchange Saturation Transfer (CEST) imaging is an emerging molecular MRI method. Amide Proton Transfer (APT) imaging, a type of the CEST-based MRI technique, is based on the chemical exchange between the amide protons of endogenous mobile protein and peptides. It has been difficult to adopt this application into the clinical routine since APT imaging is required relatively long imaging time due to multiple saturation offset image frames. Here, we propose a novel highly accelerated 4D CEST imaging technique by combining parallel imaging and compressed sensing.


Chemical exchange saturation transfer (CEST) is a novel MRI contrast enhancement technique that allows indirect detection of metabolites with exchangeable protons. The CEST effects of Amide Protons (-NH) were first introduced(Zhou 2003), and APT imaging has been widespread for a range of applications and could differentiates the brain tumor regions from peritumoral edema and normal appearing white matter(Zhou 2008). However, this CEST imaging technique in the clinical routine has been relatively limited by long acquisition time due to use multiple RF saturation offsets, even perform 3D imaging for increasing SNRs. In this work, we propose a novel highly accelerated 4D CEST imaging technique by combining parallel imaging and compressed sensing and demonstrate the feasibility of this method (up to 4-fold) in figure 4.


The 4D CEST data sets were reconstructed by using the SparseSENSE signal model(Liu 2008)(citation not found: King_K_2008) with varying RF saturate offset and applying three \({\ell 1}\)-norm penalties. As a vector form, images (m) were reconstructed by minimizing a nonlinear conjugate gradient of combining compressed sensing and sensitivity encode as the equation(Lebel 2013) below,

\begin{align} x=argmin_{m}\|\tilde{F}_{Ω}\tilde{S}m-y\|_{\ell_{2}}^{2}+λ_{1}\|Vm\|_{\ell_{1}}+λ_{2}\|Ψm\|_{\ell_{1}}+λ_{3}\|T_{v}m\|_{\ell_{1}}\notag \\ \end{align}

where \(\tilde{F}_{Ω}=[\tilde{F_{0}},\ldots,0;0,\dots,\tilde{F}_{n_{c}}-1]\in{C}^{n_{f}n_{c}n_{k}\times n_{f}n_{c}n_{v}}\) is the undersampled Fourier operator, \(\tilde{S}=[\tilde{S}_{0},\ldots,0;0,\ldots,\tilde{S}_{n_{c}-1}]\in{C}^{n_{f}n_{c}n_{v}\times n_{f}n_{v}}\) is the sensitivity matrix, \(y=[y_{0},\ldots,y_{n_{c}-1}]^{T}\in{C}^{n_{f}n_{c}n_{k}}\) is the acquired 4-D k-space data, V is a 1D spectral high-pass filter along the saturation frequency dimension and applied a symmetric window given by [0.06 0.44, 1, 0.44, 0.06], \(\Psi\) is a 4-D wavelet transform, and \(T_{v}\) is a 3-D spatial finite-differences transform. The regularization factors (\(λ_{1,2,3}\)) were set to \(λ_{1}=6.5\times 10^{-3},λ_{2}=5.7\times 10^{-3}\), and \(λ_{3}=5.1\times 10^{-3}\). These three \({\ell 1}\)-norm constraints on the data consistency term were to overcome incoherent aliasing and synthesize missing data. The 3D sensitivity maps for each 3D image of a 4D CEST dataset were separately estimated from a fully sampled region in the center of 3D k-space using a 3D extension of the eigenvalue problem approach to sensitivity maps (calibration region: [13, 6, 4], kernel size: [7, 2, 2])(Uecker 2013). An image vector \(m\in{C}^{n_{v}}\) of \(n_{v}=nv_{y}\times nv_{x}\times nv_{z}(row\times column\times slice)\) voxels denotes a vector representation of a 3D discrete image acquired at the \(v^{th}\) frequency offset. The 21 dynamic CEST data were retrospectively undersampled by modifying the GOlden-angle Cartesian Randomized Time-resolved (GOCART)(Zhu 2016) 3D MRI based on golden angle cartesian radial spoke with the reduction factors (R = 2, 3, and 4) along the RF saturation frequency offset direction. The study was performed on a healthy volunteer with a 8-channel phased array receive coil and acquired on a 3T philips Achieva system. A fully sampled CEST imaging data was acquired with a 3D multi-shot turbo field echo (TFE)(Tietze 2013)(Dula 2012) with TFE factor/echo time/flip angle(α) = 11/ 3.7 ms/ 10°, SENSE parallel imaging (reduction factor of 1), and a 1-3-3-1 binomial pulse for fat suppression and the offset frequency of irradiation was swept from -5 to 5 at intervals of 0.5 ppm with a saturation power of 0.8 µT, and plus an unsaturation pulse for CEST normalization. The total scan time for acquiring whole-brain with 16 slices was 22 min. 27 s.

Results and Discussions

In figure 2, the difference images were