Simultaneous Artifacts Correction And Acceleration For Chemical Exchange Saturation Transfer Imaging Via Deep Learning

Chemical exchange saturation transfer (CEST) magnetic resonance imaging (MRI) has emerged as a novel technology for precise diagnosis of various diseases using either endogenous molecules or exogenous administered contrast agents. CEST MRI methods rely on molecular signal saturation through radiofrequency pulses with the same frequency as a prescribed molecule (e.g., glutamate) to be measured in a magnetic field. Because the background signal from water modular is generally greater than the signal from the prescribed molecular by several orders of magnitude, there are many different types of molecules with similar magnetic frequency to be applied in CEST MRI. However, any offset from the desired frequency will produce large discrepancies due to the reduction of the background signal saturation. Current practice relies on acquiring extensive data at intentionally changed saturation frequencies and estimating the desired signal through data interpolation. This method is effective, but takes long acquisition time, making it impractical for clinical applications. To address this challenge, I developed CEST MRI methods to achieve the following two goals: 1) to accelerate the acquisition time; and 2) to increase the CEST contrast quantification quality. As the in-vivo MR environment is highly complicated and difficult to model accurately, I proposed to use deep learning (DL) to achieve these two goals. Three different methods are proposed to improve the procedure of CEST MRI: 1) a deep learning-based Glutamate CEST imaging B0-inhomogeneity correction method (DL-B0GluCEST) to accelerate the total scan time; 2) an improved DL-B0GluCEST method using data acquired from the downfield Z-spectrum only; 3) two deep learning-based methods for estimating B0 inhomogeneities from fewer calibration data. In contrast to currently practiced methods, my CEST methods are believing to be the first-of-its-kind CEST MRI methods utilizing deep learning approaches. More importantly, in my demonstrated applications, three proposed deep learning-based methods showed CEST contrast quantification quality improvement while significantly reducing CEST acquisition time by over 60%, 80%, and 80%, respectively.