The routine clinical application of CEST MRI is limited by its relatively long scan times, since CEST typically requires images to be repeat acquisitions with different saturation frequencies. Despite recent advances in fast CEST acquisition methods (1-4), the only widespread option available on MRI scanners for accelerating CEST acquisitions, is currently parallel imaging, such as with SENSE (5). However, the acceleration factor offered by SENSE for CEST MRI rarely goes beyond two (6-8). Here, a novel acquisition and reconstruction method, variably-accelerated sensitivity encoding (vSENSE), is proposed for faster acquisition than conventional SENSE. The vSENSE method undersamples k-space variably for images acquired at different saturation frequencies to generate accurate sensitivity maps and maximize acquisition speed.

The vSENSE method was validated on a doped water phantom and 8 consented brain tumor patients studied on a 3T Philips MRI scanner with a 32-channel head coil. CEST contrast was achieved using 0.8s saturation duration and 2μT saturation power (saturation frequencies, 14 to -8ppm stepped at 0.5ppm, in addition to an unsaturated reference–S₀; two-dimensional turbo-spin-echo, TSE, for image readout; FOV=212×185×4.4mm; acquisition resolution=2.2×2.2×4.4mm; acquisition matrix size=96×84x1; total duration=2.7min) (6). A 2D TSE “WASSR” (9) sequence was acquired separately for B₀ inhomogeneity correction. The raw k-space CEST data and the vendor’s preset sensitivity reference scan, were saved for offline in-house reconstruction. Conventional T₂-weighted, FLAIR, and T₁-weighted anatomical MRI were also acquired from each patient.

First, standard reconstruction was performed with the following steps: (A1) Fourier transforming (FT) the k-space data from each channel; and (A2) combining images from all channels using the “maximized SNR” strategy (10).

Second, SENSE reconstruction was performed by: (B1) computing sensitivity maps from the vendor’s preset reference scan; (B2) undersampling by increasing the phase-encoding gradient step size by 4-fold in every dynamic CEST dataset, resulting in folded images; and (B3) reconstructing unfolded images using the SENSE method (5).

Third, vSENSE reconstruction was performed by: (C1) selecting one of the CEST dynamic images to retain the full k-space data (acceleration, R=1). Here the S₀ dynamic was chosen. (C2) Reconstruct images from this fully acquired data set using FT, and fit the sensitivity maps with a weighted polynomial. (C3) Increase the phase-encoding gradient step size to variably undersample the other CEST images, with R=2 for the ±3.5ppm images and R=4 for the rest. (C4) Apply the estimated sensitivity maps to the variably undersampled datasets to reconstruct images.

Fig. 1 shows the conventional SENSE image (Fig. 1b, R=4) has a much greater difference (Fig. 1d) from the standard FT image (Fig. 1a, R=1), compared to the difference (Fig. 1e) with the vSENSE method (Fig. 1c, R=4). The vSENSE method self-estimates the sensitivity maps from the fully-acquired S₀ dynamic, resulting in more accurate estimation of sensitivity maps than conventional SENSE with a separate reference sensitivity scan. This is because the vendor’s preset SENSE reference scan acquired at the beginning of each exam virtually always uses a different acquisition sequence, resolution and orientation vs. subsequent imaging scans.

Fig. 2 from a brain tumor patient, shows conventional SENSE (R=4) yielded substantially corrupted Amide Proton Transfer (11) weighted (APTw, part e) and raw (part g) images, compared to the standard FT results (blue arrows indicate tumor region, and red arrows indicate SENSE-reconstruction artifacts). In contrast, the vSENSE (R=4) method generated APTw (part f) and raw (part h) images, as well as z-spectra (part i) highly consistent with FT ones. Note that the sampling pattern can be adapted for other CEST imaging applications, such as creatine (12) or glucose(13, 14) imaging.

The vSENSE method variably undersamples saturated images at different frequencies, resulting in an overall acceleration factor of essentially four. In conjunction with the fully-sampled S₀ dynamic, the vSENSE method guarantees accurate sensitivity maps and sufficient SNR for accurate APTw imaging. As implemented here, vSENSE doubled the speed and provided better results than conventional SENSE, which may prove key for advancing CEST to the clinic.