Introduction

Chemical exchange saturation transfer (CEST)-MRI is emerging as a method for in vivo imaging of metabolite levels in intact tissue. Cardiac CEST-MRI has been used to probe creatine levels in the setting of myocardial infarction^1,2 and obesity³, and identified reduced myocardial creatine CEST-contrast in failing tissues. When imaging the heart, the need to time saturation pulses and acquisition to the cardiac cycle results in significantly longer scan times than stationary organs. Subsequently, CEST-MRI is typically performed by acquiring fewer images along a z-spectrum and calculating CEST contrast via image subtraction instead of z-spectral fitting. While accelerated imaging via compressed sensing reconstruction has been explored in several dynamic contrast settings, incorporation of compressed sensing for accelerated image acquisition has not been fully explored in the setting of CEST. In this study we probed how reconstruction of increasingly under-sampled CEST-MRI data using various compressed sensing methods inserted bias into resultant z-spectra and measured CEST contrasts.

Methods

One C57BL/6 mouse was anesthetized with 1 – 3% isoflurane gas and imaged using a 7T MR scanner (PharmaScan, Bruker, Ettlingen, Germany) and a 2 × 2 surface array coil (Bruker). A custom sequence was designed such that CEST-preparation (4 Gaussian shaped pulses, B1 = 0.4 μT, total sat time = 604 ms) was ECG triggered and a 4 golden-angle incremented radial acquisitions were performed during diastole (TR/TE = 30 / 2.1 ms, 75 segments, matrix = 192 × 192, FOV = 25 mm × 25 mm, slice thickness = 1 mm, NA = 1). The total sat time and power were intentionally chosen to generate minimal creatine and APT contrast in resultant z-spectra. In one short axis midventricular slice a reference image (saturation offset = 30ppm, B1 = 0.4μT) and four spectral data sets were acquired at fully-sampled, 2x, 4x and 5x-under-sampled. Spectral data sets contained 101 saturation offsets from -10 to 10 ppm with step size of 0.2 ppm. Afterwards, images were reconstructed with seven different compressed sensing methods using the BART Toolbox⁴. Nonuniform fast Fourier transform (NUFFT) reconstruction of the fully sampled data was used as a ground truth, while under-sampled data was reconstructed with 3D isotropic total variation, ℓ1-norm, ℓ2-norm, ℓ1 wavelet transform, total variation, and locally low rank reconstructions with regularization parameter = 0.001 algorithms. A region of interest in the septum was defined on an NUFFT image, applied to all image sets, and the subsequent z-spectra were fit using the sum of 5 Lorentzian functions representing water, creatine, amide proton transfer (APT), NOE effect, and magnetization transfer (MT).

Results

Representative CEST encoded images for fully sampled and 2x under-sampled data are shown in Figure 1 and represent the range of reconstructed images at offsets relevant for creatine CEST-MRI. All reconstructions yielded images with poorer definition of cardiac anatomy, however ℓ2-norm reconstruction demonstrated substantially lower signal to noise in the heart compared to other methods (Figure 1). Similar results were observed at 5x under-sampling as shown in Figure 2. However, reconstructed Z-spectra from 2x under-sampled data demonstrated similar and substantial bias regardless of which method was used across the z-spectrum (Figure 3). This was not observed at higher levels of under-sampling (Figure 3). Creatine and APT CEST contrasts calculated from fully-sampled data were 0.4% and -3.0%, respectively. Creatine CEST contrast was uniformly higher across compressed sensing algorithms when 2x (11.0 ± 0.4%) and to a lesser extent 4x (7.7 ± 0.1%) under-sampled data were reconstructed (Figure 4). APT CEST contrast calculated from 2x under-sampled data underestimated fully sampled data (-5.8 ± 0.1%), however increased under-sampling resulted in greater positive bias of calculated APT-contrast (Figure 4).

Discussion

The results of this study suggest that simple application of compressed sensing protocols to cardiac CEST data may result in biased measurement of common endogenous CEST contrasts. For measurement of myocardial creatine in particular, significant positive bias may result at low rates of undersampling across compressed sensing algorithms. Further understanding of the mechanisms behind changes in signal intensity due to compressed sensing reconstruction of spectral image sets is necessary.

Acknowledgements

This study was supported by NIH 1R01HL28592.