Simultaneous mapping of B0, B1 and T1 for the correction of CEST-MRI contrast
Final Report Abstract
With this project, we successfully demonstrated the feasibility of simultaneous mapping of the static magnetic field (B0), the RF excitation field (B1) and the longitudinal relaxation time (T1) using a modified WASABI MRI sequence in combination with a neural network-based analysis. Using a single MRI sequence for all three parameters instead of separate scans for each parameter simplifies the data acquisition and reduces the acquisition time. Furthermore, once trained, the NN-based analysis reduces the post-processing time compared to conventional fitting approaches using least squares optimization. For the simulation of WASABI and CEST spectra, two different simulation tools have been implemented that both solve the matrix equation representation of the Bloch McConnell (BMC) equations numerically. While the BMCTool is more flexible due to its implementation in Python, the PyPulseq- CEST tool provides a higher performance due the partial implementation in C++. Both tools were used for this project as well as for the Pulseq-CEST framework and the resulting publication. To increase the sensitivity of the modified WASABI approach towards the T1 relaxation time, two major changes were made in the modified compared to the original WASABI pulse sequence. First, a T1 preparation block was introduced that saturates all remaining magnetization of the previous MRI readout and thus leads to a higher range of signal changes and ensures identical starting conditions at every frequency offset. Second, the previously fixed trec delay was made offset-dependent to cover broader T1 ranges. A major part of the project was the implementation and optimization of the NN that was used for the parameter prediction and uncertainty estimation. A Gaussian negative log likelihood (GNLL) loss function was chosen for the supervised training on simulated data generated using the developed simulation tools. The NN-based analysis was found to be superior compared to conventional fitting approaches especially in cases where the initial parameter estimates/starting values of the fit deviated from the true values. The uncertainty estimates of the NN were found to correlate well with the observed differences between the predicted parameters and ground truth values. Consequently, the standard deviation of the assumed Gaussian distribution was shown to be an appropriate measure for the uncertainty of the proposed approach and thus enables the identification of cases or regions where the predicted parameter values should be treated with caution. This is of high importance when the proposed approach is transferred to clinical setups where the generated parameter maps will be used for the correction of CEST-MRI contrasts. Parts of this work have been presented at the 2021 ISMRM & SMRT Annual Meeting & Exhibition and at the 23rd Annual Meeting of the German Chapter of ISMRM in Zurich, where it was awarded the price for the best scientific poster. An article summarizing all results from this project is currently in preparation.
Publications
-
Pulseq‐CEST: Towards multi‐site multi‐vendor compatibility and reproducibility of CEST experiments using an open‐source sequence standard. Magnetic Resonance in Medicine, 86(4), 1845-1858.
Herz, Kai; Mueller, Sebastian; Perlman, Or; Zaitsev, Maxim; Knutsson, Linda; Sun, Phillip Zhe; Zhou, Jinyuan; van Zijl, Peter; Heinecke, Kerstin; Schuenke, Patrick; Farrar, Christian T.; Schmidt, Manuel; Dörfler, Arnd; Scheffler, Klaus & Zaiss, Moritz
