Chemical exchange saturation transfer (CEST) imaging is a promising molecular MRI technique in which the contrast is caused by a change in water signal intensity due to chemical exchange with saturated solute protons (1). CEST imaging can detect many low-concentration biomolecules in tissue, such as proteins with amide proton transfer (APT) imaging (2), glycosaminoglycan using GagCEST (3), and glutamate using GluCEST (4), and has much potential in the clinic, such as detecting cancer (5-7), stroke (8,9), and Parkinson's disease (10). However, CEST MRI typically requires relatively long scan times, due to the use of multiple RF saturation frequencies, possible multiple acquisitions to increase SNRs, and a long RF saturation pulse (or multi-pulse train), thus limiting its clinical translation. Herein, as proof of principle, we demonstrated the feasibility of an ultrafast APT imaging (up to 10-fold) based on blind compressed sensing (BCS), which models the dynamic CEST signal as a weighted linear combination of temporal basis functions from a dictionary learned from the undersampled measurements (Fig. 1).

Dynamic CEST signal recovery, γ(x, w)=∑U*(x)V*(w) was formulated as a constrained optimization problem (11):

$$$[U^{*},V^{*} ]=argmin[||A(UV-b)||_F^2+\lambda_{1}||U||_{1}],s.t. ||V||_F^2<1$$$

where the criterion is a linear combination of data consistency term (b=undersampled data, A=undersampled Fourier encoding matrix with coil sensitivity) and sparsity-promoting l1 norm prior on spatial coefficient matrix U, subject to the Frobenius norm constraint on the dictionary V. The BCS model simultaneously estimates the sparse spatial coefficients U and the dictionary basis functions V from the undersampled k-space measurement b. To demonstrate the ability of the BCS algorithm, the computer model was implemented to mimic the APT experiment. The simulated phantom contains 9 compartments with different APT-pool concentrations, exchange rates, and semi-solid MT concentrations (water: w, APT: s, MT: m, B₀/B₁=3T/2μT, T_sat=5s, T_1w/T_2w=1.4/0.1s, T_1s/T_2s=1.4/0.05s, T_1m/T_2m=1.4s/20μs). For the in-vivo study, all healthy volunteers (N=2) and patients (N=5) were scanned on a Philips 3 T MRI scanner after informed consent was obtained in accordance with the IRB requirement. A fully sampled 2D CEST saturated images were acquired using a TSE sequence with an RF saturation time of 800ms, RF saturation frequency offsets (14 to -8 ppm at intervals of 0.5 ppm), and varied RF saturation power levels (0.5, 1, 1.5, 2 μT for healthy volunteers and 2 μT for tumor patients). The dynamic CEST data was retrospectively undersampled by the golden-angle ordering pattern along RF saturation frequency direction at acceleration R=2, 4, 10, 20, and 50. Water saturation shift-referencing (WASSR) data (26 frequency offsets from 1.2 to -1.2 ppm at intervals of 0.125 ppm and a saturation power of 0.5μT) was acquired for B₀ corrections (12). As previously used in (2), APT-weighted (APTw) images were quantified by MTR_asym(3.5ppm).

Fig. 2 showed the BCS reconstructed APT images of the simulated phantom and the Z-spectrum comparison to the Nyquist sampled, low-resolution acquisition scheme. Accurate APT imaging was successfully achieved using the BCS reconstruction with acceleration factors up to 10. Fig. 3a compared brain APTw images of a typical healthy volunteer (with four different RF saturation powers) under full k-space and BCS acceleration factors of 10. Importantly, accurate BCS reconstructed APTw images without compromising image quality were achieved. The correlation and Bland-Altman analysis were performed to assess the bias and limit of agreement between the full k-space and BCS reconstructed APTw images (Fig. 3b). The BCS reconstructed APTw image had a high correlation with the full k-space reference image (r²=0.95 for R=10), and the Bland-Altman analysis showed that the bias was significantly small (mean difference of 0.0019% and absolute values for the limits of agreement (±1.96 SD) of +0.69% ~ -0.72%). In the tumor patient study, the Gd-enhancing area (tumor core) on the T₁-weighted post-Gd image was hyperintense on the APTw image (Fig. 4). Again, the accurate APTw image was successfully reconstructed using the BCS algorithm with acceleration factors of up to 10-fold. The mean difference of APTw signals between the full k-space reference and BCS reconstructed images was only 0.15% and 0.09% for the normal (normal appearing white matter) and tumor region (Gd-enhanced tumor core), respectively.We have proposed and valuated a novel BCS-based, dynamic CEST MRI technique for clinical applications. Simulation and in-vivo experiments demonstrated the feasibility of such ultrafast CEST imaging with an acceleration factor up to 10. BCS-based APT imaging acquisition can significantly reduce the scan time and potentially has a wide range of clinical applications, particularly for pediatric patients.