Scientists or clinicians interested in fast CEST imaging.Chemical exchange saturation transfer (CEST) imaging is a promising novel technique and many promising applications have been proposed, such as cancer and stroke evaluation [1]. CEST requires acquisition of saturation images at multiple frequencies, the so-called z-spectrum, which can be time consuming, thus hampering its clinical translation. With the recent emergence of compressed sensing (CS) theory [2,3], several techniques have been developed applying CS in MRI [4,5]. Blind compressed sensing (BCS) [6] is a novel method, which represents the dynamic signal as a sparse linear combination of temporal basis functions from a large dictionary. In this work, we replace the original temporal dependence with off-resonance saturation frequency dependence and propose a sparse dictionary learning algorithm utilizing the spatial-frequency prior information in CEST imaging. Here we provide a proof-of-concept study in accelerating CEST imaging using this algorithm.The central idea of BCS is to represent the dynamic image series as the product of the coefficient matrix U and temporal/frequency basis functions matrix V. Assuming only few frequency basis functions are enough to describe the dynamic behavior of all pixels, creates a constraint on the sparsity of the coefficient matrix U. In CEST imaging, voxels in the same compartment have similar z-spectra [7], which show group sparsity in the spatial-temporal domain. These z-spectra can be treated as temporal/frequency basis functions in the dictionary. The original BCS method addressed the sparsity in the U matrix, without considering the group sparsity in spatial-temporal domain. To address the group sparsity in the spatial-temporal/frequency domain in CEST imaging, we impose the sparsity of coefficient matrix in the wavelet and total variation domain to better utilize the prior information in CEST imaging. The reconstruction algorithm can be expressed as: $$$\min_{U,V}\|F(UV)-d\|_{2}^{2}+\alpha\|U\|_{1}+\beta\| \Psi (U)\|_{1}+\gamma TV(U)+ \eta \|V\|_{2}^{2}$$$, where F is the undersampling Fourier operator, Ψ is the wavelet operator, TV is the total variation operator, d is the undersampled k-space data. This model coincides with the formulation of Sparse BLIP [8], and we solve it with a similar alternating optimization technique. Measurements were performed on a Philips 3T Ingenia system using a 32 channel head coil. The phantom consists of 5 tubes of iopamidol solution with pH values of 6.0, 6.5, 7.0, 7.5 and 8.0 within a water-filled container. The CEST images were acquired with a TFE sequence, flip angle of 45°, TR/TE=4.5/2.0 ms, slice thickness=10 mm, matrix=128x128, FOV=256x256 mm. Saturation RF consisted of a train of 92 Sinc RF pulses, each of 21.5 ms duration with 0.5 ms pulse intervals, swept between ±800 Hz in steps of 38Hz. CEST processing used standard methods [9] with WASSR correction for $$$B_{0}$$$ inhomogeneity. We evaluated the performance of the BCS scheme by performing retrospective undersampling using Cartesian sampling and radial sampling. Cartesian trajectory schemes use variable density undersampling pattern [4] with 16 k-space lines, acceleration factor R=8. Radial trajectory schemes use continuous golden angle rotations (111.25° angular increment) [10,11] with 12 spokes, acceleration factor R=12. Both of them are incoherent sampling patterns satisfying the CS requirement.Fig. 1 compares CEST effects reconstructed with fully sampled, BCS, and the proposed method from Cartesian sampling and radial sampling. The CEST effects are measured at 537 Hz (4.2 ppm). Fig. 2 (a-b) compares z-spectra reconstructed with fully sampled, BCS, and the proposed method from Cartesian sampling and radial sampling in each region of interest (ROI) for different pH values. The proposed method performs better than BCS for both of the sampling patterns. Since the CEST effect is relatively weak, it is sensitive to the aliasing and artifacts in spatial-temporal domain. The group sparsity prior information in spatial-temporal domain is helpful to improve the reconstruction quality at high acceleration factors.We propose a spatial-temporal sparse dictionary learning algorithm, which is an enhanced version of the blind compressive sensing method to reconstruct the image for undersampled CEST data. Phantom results demonstrate that the proposed method is able to improve CEST imaging at high reduction factors for both Cartesian and radial sampling. The specific acceleration times would depend on the details of experimental implementation. Work is underway to experimentally test accelerating in vivo CEST imaging using the undersampling pattern and reconstruction algorithm described in this study.