DeepCEST: 9.4 T Chemical Exchange Saturation Transfer MRI Contrast Predicted from 3 T Data - a Proof of Concept Study

Results of the deepCEST network compared to Lorentzian fitting in the training dataset. Real 9.4 T CEST fits (column 1), prediction of net (column 2), prediction difference from data (column 3). ROI evaluations for these data can be found in the Supporting Information.
Chemical exchange saturation transfer (CEST) allows for indirect detection of solute molecules via exchanging protons that transfer selectively applied saturation to the large water pool in tissue. While studies have been performed at clinical field strengths, CEST effects can be studied more specifically at higher field strengths where peak separation and selective saturation benefit from the increased frequency separation between resonances, proportional to the Larmor frequency. Some of the peaks can be detected separately only at UHF and lead to the understanding that some signals detected at 3 T are actually mixed signals from several resonances. The information is not gone, but just hard to extract, and 3 T signals are still rich in information from different origins.This work follows the approach of using prior UHF knowledge for 3 T evaluation, in this case by applying artificial neural networks to combine these different data. The proposed neural network is trained using 3 T Z‐spectra as an input and 9.4 T CEST parameters as a target. Thus, it is trained to predict 9.4 T CEST contrasts from a Z‐spectrum measured at 3 T. In a way, this is the most direct approach of using 9.4 T prior knowledge for 3 T data evaluation. While application of neural networks in the field of MR has gained more interest in recent years, the presented approach represents a first step toward application in CEST MRI and is a rather simple approach. A multilayer perceptron is used to combine coregistered data acquired at a 9.4 T human MRI scanner and a 3 T clinical scanner.