CEST analysis using a deep residual learning method based on Bloch-McConnell synthetic training data

Motivation: Conventional Lorentzian-based fitting methods for CEST MRI are time-consuming and susceptible to initial and bound values. Goal(s): We aim to propose a deep residual learning method based on Bloch-McConnell synthetic training data to quantify CEST contrasts Approach: We utilized deep residual networks to predict the Z-spectrum reference (Zref) signal (ZrefNet) for subsequent CEST quantification (Zref - Z). ZrefNet is trained purely on Bloch-McConnell synthetic data, and tested on simulation and in vivo human brain data. Results: ZrefNet generated accurate CEST contrasts rapidly with high correlation to ground-truth on simulation data and demonstrated feasibility for analyzing human multiple sclerosis / healthy control CEST data. Impact: We proposed deep residual networks to predict the Z-spectrum reference signal (ZrefNet) for CEST quantification. ZrefNet generated accurate CEST contrasts on simulation data and demonstrated its feasibility for analyzing in vivo human CEST data of multiple sclerosis and healthy controls.