ComprehenCEST: a clinically feasible CEST protocol to cover all existing CEST preparation schemes by snapshot readout and reduction of overhangs

Introduction CEST sequences preparing a selected metabolite normally sample the full Z-spectrum, allowing for asymmetry or model evaluations. To achieve good labelling efficiency for various metabolites, different physical preparations are required. Sparsity-enforcing machine learning algorithms help to select and combine the differently CEST-prepared images from a sequence pool covering the preparation parameter space, while preserving main contrast information. Together with a fast, single-shot 3D readout, we create a 3D CEST protocol containing 13 established contrasts in 10 minutes scan time. Materials & Methods 13 contrasts are evaluated conventionally (PCA-denoising, dB0-correction, MTR asymmetry or Lorentz fit) from six existing CEST sequences [2-5], which cover B1cwpe levels from 0.5 to 4 uT, offering the ground truth. Mz is prepared using standardized pulseq-CEST building blocks [7-8], and probed with the fast snapshotCEST 3D readout [1]. (2) All Z-spectra, B0 and B1 are mapped to contrast via a linear projection, while sparsity-enforcing L1-regularization reduces the number of consumed offsets (rowLASSO [6]). The training is carried out on uncorrected, raw Z-spectra to generate a selection that provides robustness against noise and B0/B1 field inhomogeneities. (3) Difference maps between ground truth and model output are created for volunteer data in the training and validation set. Results The normalized, mean absolute error (NMAE) between linear model and ground truth, averaged over all 13 contrasts, and for retaining offset rate r is: 27±6% (r=1), 29±7% (r=0.8), 32±8% (r=0.6), 36±9% (r=0.4), 42±11% (r=0.2). Discussion & Conclusion Instead of arguing which is the best CEST protocol to provide new insights into a pathology, and only measure one CEST contrast, we suggest measuring them all. By combining sparse sampling and snapshot readout, a comprehensive protocol covering most of the reported labellings of 10 minutes is conceivable. This allows to design powerful hypotheses generating clinical pilot studies