Machine Learning accelerates and stabilizes selective CEST fitting at 3T

Multi-Lorentzian analysis of chemical exchange saturation transfer (CEST) Z-spectra by non-linear least squares (NLLS) fitting is common at ultra-high field strengths but particularly challenging at clinical field strengths due to broad, coalesced peaks and low SNR. Here we demonstrate that a neural network (NN) trained on just 3 slices of a single subject can robustly predict CEST Lorentzian pool parameters in other subjects, in the presence of motion, and in a brain tumor patient, with a 95 % reduction in computing time, allowing for quick estimation of NLLS initial conditions or initial online reconstruction of spectrally selective CEST contrasts