A deep learning-driven CEST signal discovery model for transmit inhomogeneity correction with single-B1 acquisition

Motivation: Most B1 inhomogeneity correction approaches require time-consuming multiple z-spectra acquisition under several B1 levels. Goal(s): To develop an interpretable deep learning pipeline for reliable B1-inhomogeneity corrected CEST imaging without the requirement for multi-B1 acquisitions. Approach: A novel transformer-based model is designed to predict multiple CEST spectra across a wide range of B1 levels, which originally needed to be acquired during imaging. Equipped with this smart model, our pipeline can yield reliable B1 corrected CEST signals. Results: The proposed approach enables quality and robust B1 correction on clinical CEST data, providing an effective way to improve quantitative CEST imaging. Impact: Our proposed deep learning-driven signal discovery model facilitates precise B1-inhomogeneity corrected CEST imaging with single-B1 acquisition, while also offering extending applications, such as improved CEST quantification via additional predicted signals.